Denes Csaszar · Berlin · change record

What changed here.

Every published change to a stated figure or a claim on this site, newest first, with what it said before. The figures themselves are on the record and in the study, and the reasoning behind the larger ones is on the journey page; this is the flat list, at the size of the edit rather than the size of the story.

This site went live on 1 October 2026. Entries dated before that record changes made while it was in preview.

1 · What gets an entry

A stated figure, a claim, or a correction. Typography, layout, code and wording that leaves the meaning intact get none. So silence on this page means nothing changed that this site asserts, and not nothing happened.

Where this record starts, and what is not in it

It begins on 1 September 2026. Changes before that date are in the repository history and in the dated correction notes inside the dataset README, which is where they were written at the time. The two entries dated before it reach back further because they were reconstructed from those notes, not because the page was running.

Some changes get no entry on purpose, and the largest of them is worth naming once rather than leaving to the silence above. In September 2026 a pass over these pages replaced the inline retellings of earlier corrections with dated marks that link here, and cut a set of sentences that said nothing the page did not already say. No stated figure and no claim moved in any of it, and every correction those retellings told is still listed below, so none of that pass has an entry of its own. A later pass in the same month cut three repetitions: the snapshot sentence off the journey page, and the premise that answer engines are not deterministic off the journey and method pages, leaving it on the record page and the study page in one wording. Nothing stopped being claimed.

⚠ And one thing this page cannot do: it is written by the same person who makes the changes. It is a record, not a check. The published files are the only thing here that anyone else can currently check for themselves, by recomputing a figure from them.

2 · The record

31 October 2026

The first step of the plan on the record page, in English, German and Hungarian, is dated as done: the study is archived on Zenodo (doi:10.5281/zenodo.23780961) and the plain list of which numbers on this site recompute has a page of its own. The same step says that the answers of the last wave cannot be re-derived.

before: None of it has started.
after: The first step is done and dated above; none of the rest has started.

The plan said each step would be dated on the record page when it happens. The archive and the list now exist, and the archive’s own note on what is lost is repeated where the step is stated.

31 October 2026

The dates on which the questions were registered and the number of engines in the first wave were restated from the capture ledger, on the record, study, method and journey pages, in English, German and Hungarian.

before: eighteen questions frozen and registered on 4 July; the first wave asked eight engines; the error caught and corrected on the day of that wave
after: ten questions registered on 4 July and all eighteen by 9 July; the first wave asked six engines; the two wrongly written questions corrected on 5 July

The capture ledger records when each run’s questions were registered and which engines answered. It shows ten questions registered on 4 July, eighteen on 9 July, six engines in the first wave and Meta AI and Mistral from 10 August. The earlier sentences said more than the record holds. The entry of 3 September 2026 below repeats the old account and is left as it was written.

28 October 2026

Part of the 3 September change is reversed: the site names Makefield again. The record page, in English, German and Hungarian, says the author publishes this work as a person and that it is also part of Makefield, with a link to makefield.io, and the study page links its listing there as Makefield study MF-2026-001.

before: no sentence on the record page or the study page naming Makefield; the name stood only inside the capture-wave labels, such as makefield-cycle2, and on this page
after: record page: “I publish this work as a person. It is also part of Makefield, my independent research project on the same question.”; study page: “Listed as Makefield study MF-2026-001”

On 3 September the attribution line stopped naming a company, because the author publishes as a person and not as a business. That stays: the author still publishes here as a person, and the attribution line still reads “Denes Csaszar”. Since 19 October the study page’s structured data has said that the study belongs to the series of studies published on makefield.io; a reader should be able to see that too. The two corrections of 3 September that stopped pages asserting current clients stay as they are. The dataset files are unchanged.

28 October 2026

The plan on the record page, in English, German and Hungarian, now says the wider set of companies will be measured in a browser and through the providers’ published programming interfaces, not through the interfaces alone.

before: through the providers’ published programming interfaces rather than a browser
after: in a browser and through the providers’ published programming interfaces

Capture in a browser continues alongside the interfaces (decided 19 October 2026). The plan had said a browser would not be used, and a reader should find the change where the plan is stated.

19 October 2026

The record page, in English, German and Hungarian, now sets out the plan for the work beyond the paraphrase study, in section 9 and marked as a plan: freeze this study, test the floor test on simulated data and on repeats where nothing should change, separate time from sitting, and only then a wider set of companies. It says that none of it has started, and that no one independent has agreed to re-run any of it. The paragraph above it, and the method page in all three languages, now also say that the paraphrase study is on hold, as the journey page already did.

before: “still a draft awaiting ratification”
after: “still a draft awaiting ratification, and it is currently on hold because of findings from its own review” on the record page; “still a draft awaiting ratification and currently on hold because of findings from its own review” on the method page

The page already said that the next study is a plan, not a result. The plan has changed shape, and a reader should meet the change where the old plan is stated, marked as a plan. The journey page said the paraphrase study is on hold and the record and method pages did not, so a reader got a different status depending on the page.

19 October 2026

On the English and Hungarian pages, the footer now says after the links to the Impressum and the Datenschutz page that they are in German and that the German version applies. Nothing else in the footer changed.

A reader on an English or Hungarian page met two German links without being told why they were there. The German legal pages are the ones that apply; the English and Hungarian legal pages are translations of them.

19 October 2026

The study page gains a figure in section 5, after the sentence that says the within-cell result is not net of period: every pair of capture dates, each cell measured against its own shortest gap, with a caption that says what the rows show. The German and Hungarian readings carry the same figure.

The page said the result is not net of period and showed nothing that let a reader see what that means. The figure puts the pairs within a month beside the pairs across the July to August boundary, and its caption says what it cannot do: tell the month from the gap. It adds a descriptive claim, so it gets an entry. Every number in it is computed from the published gap-pairs file when the page is built, and the build fails if the drawing and the file disagree.

19 October 2026

The study page’s structured data now says the study is also part of a series published on makefield.io. Nothing a reader sees on the page changed.

before: isPartOf: this website
after: isPartOf: this website, and https://makefield.io/studies/#series

The study is MF-2026-001 in the Makefield studies series, which makefield.io declares and publishes. The study, its author and its dataset are still declared here and only here; the series is the one thing that lives there, so this page points at it rather than describing it a second time.

19 October 2026

The structured data that describes the author now names two profiles of the same person: his LinkedIn profile and his ORCID record. Nothing a reader sees on the page changed.

before: sameAs: none
after: sameAs: linkedin.com/in/denescsaszar, orcid.org/0009-0002-1002-1609

Until now the markup listed sameAs as left out on purpose, because no verified profile existed to point at. Both now exist and belong to the author. A link to Makefield is not among them: an organisation is not the person.

19 October 2026

The footer of the German and Hungarian pages no longer links the English legal notice and privacy notice. German pages link the German ones, which apply; Hungarian pages link the German ones and their Hungarian translations. Every English page, and the German legal pages themselves, still link the English ones.

The 1 October entry says the English legal pages are linked from the footer of every page. On a page read in German or Hungarian they repeated, in English, links already there in the language that applies or the reader's own. That entry stays as written; this one narrows it.

19 October 2026

Ten sentences on the record page, the method page, the journey and the study page said more, or something other, than they meant, in English, German and Hungarian. Blind Hungarian readers stumbled on each of them while translating, and in each case the Hungarian said what the English said.

before: “p below the floor of the test”; “The wave this figure comes from”; “because priming is the exact bias the measurement exists to detect”; “and one of the eleven carries no date at all”; “for that failure to pull a published number” and “Whether it pulls a published number”; “a confound of that kind pushes the share of cells that moved up rather than down” and “where it inflates the reading rather than damping it”; “A threshold set without one can sit below the rate a working instrument produces”; “It is net of engine and question, not net of period”; “because a score formed mid-run is a score formed while the next question is still unasked”; “Nothing was published and nothing was claimed on it”
after: “p below the smallest value 5,000 replicates can show”; “The waves this figure compares”; “because priming would create exactly the bias the measurement exists to detect”; “and one of those eight carries no date at all”; “to withdraw a published number” and “Whether it affects a published number”; “if a confound of that kind acts at all, it can only push the share of cells that moved up, not down” and “where, if it acts at all, it can only inflate the reading, not damp it”; “can sit so low that chance alone crosses it”; “It is net of engine and, except for two cells named below, of question; it is not net of period”; “because forming a judgement before the run is complete could influence how the remaining questions are captured”; “Nothing had been published and nothing claimed on it at that point”

No figure moved. “Floor” already means the noise floor everywhere else here. The figure compares waves and comes from none of them. Priming is excluded as a control; nothing measured it. Three dates, eight unpublished and one undated read as twelve files where there are eleven. “Pull” meant withdraw in one sentence and affect in the next, on a page that then says nothing is withdrawn. Sign-in state was never recorded, so the pages may say which way it would push, not that it did. The within-cell result rules out an engine composition effect and, for all but two cells, a question one; the first sentence now says so before the paragraph that names them. The registration sentence is about 22 August, and read today the old tense said the floor test is still unpublished.

19 October 2026

The privacy notice gave 10 September 2026 as its date, in German and in its translations. It now gives 22 September 2026.

before: „Diese Erklärung hat den Stand 10. September 2026.“
after: „Diese Erklärung hat den Stand 22. September 2026.“

Section 3 was changed on 22 September, when the account's exact creation time replaced the upper bound, and it names a value read on 17 September. A notice dated before its own last change tells a reader its date cannot be trusted, which is the argument of the 1 September entry about revision dates. The English and Hungarian translations carry the German date and move with it.

19 October 2026

The legal notice and the privacy notice now also exist in Hungarian, at /impresszum and /adatkezeles, linked from the footer of every Hungarian page beside the German ones.

A Hungarian reader should not need German or English to learn who runs this site and what it processes. The German pages remain the ones that apply; each Hungarian page says so in its first line and states nothing the German page does not.

19 October 2026

A statement on how the captures were obtained came off the record page, the study page and the method page, in English, German and Hungarian, and out of the dataset README, together with the parts of this record that retold it. No figure changed.

It said nothing about any figure on this site, and every figure is unchanged. What a reader needs to judge them, one account, sign-in not recorded and the capture mode fixed per engine, stays on the record, study and method pages.

3 October 2026

The record page, the study page, the method page and the journey now also exist in Hungarian, at /hu, /hu/drift-2026-08, /hu/modszer and /hu/ut, each linked to its English page and back.

A Hungarian reader should not need English to check this work. Every figure on the Hungarian pages is read from the same source as the English ones and the build fails if one differs. Where the readings differ, the English pages are the reference. This change record stays in English; every link to it from a Hungarian page says so.

1 October 2026

Four pages said the floor test was written on 22 August “after the data existed”. Six of the seven waves had been captured by then; the seventh was first captured on 25 August.

before: “the floor test itself was written on 22 August, after the data existed” (record page, and in the same words on the study page and the method page); “The analysis was written after the data existed.” (journey)
after: “… after six of the seven waves had been captured”; “The analysis was written after six of the seven waves were captured.”

The wave list on the study page dates the last wave’s first capture to 25 August, three days after the analysis was written. The sentence was meant as a limitation and it still is one, but it claimed more than the dates support. Found by the Hungarian translation’s build seat while checking the dates.

1 October 2026

The record page, the study page, the method page and the journey now also exist in German, at /de, /de/drift-2026-08, /de/methode and /de/weg, each linked to its English page and back.

A German reader should not need English to check this work. Every figure on the German pages is read from the same source as the English ones and the build fails if one differs. This change record stays in English; every link to it from a German page says so.

1 October 2026

The record page and the study page said an unrecorded factor between the waves pushes “the reading” up, without saying which quantity. It is the share of cells counted as moved.

before: “where a confound of that kind pushes the reading up rather than down” (record page); “bias the reading upward” (study page)
after: “pushes the share of cells that moved up rather than down”; “bias the share of cells that moved upward”

“The reading” could mean the share, the median difference or the result as a whole. The sentence is about the count: a factor that makes answers less alike between waves makes more cells fall below their own floor. Three independent readers of the German translation could not tell which quantity was meant, and the English gave them nothing more to go on.

1 October 2026

The download block, on the record page and the study page, and section 6 of the study page said four things are read out of the summary file rather than checked against it. Two more are: the number of records folded into cells and the number of draws dropped because the engine returned no answer. Neither CSV carries them.

before: “four things are read out of it rather than checked against it: the permutation null, the twelve-cell coverage check, the two bootstrapped intervals and the count of excluded cells”
after: “six things are read out of it rather than checked against it: …, the count of excluded cells, how many records were folded into cells and how many draws were dropped because the engine returned no answer”

A list that says which figures a reader cannot check is only worth something if it is complete. Found while building the German study page, when both counts had to be marked as coming from the summary because there was nothing else to recompute them from.

1 October 2026

The record page justified dropping a pair whose two token sets are both empty with draws that “cited nothing”. The comparison runs over the content tokens of the answer text, not over citations.

before: “because two draws that cited nothing are not evidence of agreement”
after: “because two draws with no content tokens left are not evidence of agreement”

The rule is right and the reason given for it described a different measurement. An empty set here means that no content word survived the filtering, and the sentence now names what is actually empty. Found by the blind German reader while translating the page.

1 October 2026

The journey page said every share is now reported by rank block of the floor. The headline share is still one pooled figure, which the study page calls an average over the four blocks.

before: “Every share is now reported by rank block of the floor”
after: “The share is now also reported by rank block of the floor”

“Every” included the one share that is not split, and it is the one a reader meets first. The split stands beside the headline; it does not replace it.

1 October 2026

The journey page said a third of the baselines rest on a single observation. A floor is measured from a pair of answers, and the record page already said a third of the floors rest on a single repeat pair.

before: “A third of the baselines rest on a single observation.”
after: “A third of the floors rest on a single repeat pair.”

One observation cannot give a floor at all, so the sentence described a measurement that does not exist here. The record page’s wording was the accurate one and the two pages now say the same thing.

1 October 2026

The journey page and the study page still called the one company the questions are about a subject the engines know little about, a claim the record page dropped earlier the same day.

before: “And the subject is close to unknown to these engines by design” (journey page); “this one runs on a low-information subject by design” (study page)
after: “And the subject was chosen because little has been published about it”; “this one runs on a subject chosen because little has been published about it”

Nothing on this site measured what the engines know about the subject, which is why the record page was corrected. That correction named only the record page, and the same claim stood on two others. A retraction that holds on one route and not on the next leaves a reader two answers to the same question.

1 October 2026

The journey page and the study page still said that no second person had scored or rated anything, the sentence the record page replaced on 17 September.

before: “No second person has scored anything.” (journey page); “No second person has rated anything.” (study page)
after: “Every number comes from one pipeline, written by the person reporting it and run over an archive captured by the same person.” on the journey page; on the study page the sentence is gone, because that pipeline sentence already follows it

Nothing here is scored by a human rater, so the sentence implied a rating step that does not exist. That is the reason given in the entry of 17 September below, and that entry changed the record page only.

1 October 2026

A sentence on the journey page about the headline and its better-measured subset said the reverse of what it meant.

before: “a reader given the headline alone would discount the subset on finding it later”
after: “a reader given the headline alone would discount the headline on finding the subset later”

The paragraph argues for printing the subset beside the headline. A reader who meets the subset later trusts the headline less, not the subset, and the sentence had the two the wrong way round.

1 October 2026

The record page now opens with the question the work answers, groups its sections under five named parts, and closes with a section on what the result is worth, what it leaves open, and the next study on the instrument.

A reader met the floor test before knowing what it was for, and the page was so careful about what the result does not show that it never said what it does. The new section says it in one sentence, scoped to this archive, and names the next study with its status: a draft awaiting ratification, a plan and not a result.

1 October 2026

The legal notice and the privacy notice now also exist in English, at /legal-notice and /privacy, linked from the footer of every page beside the German ones.

The site is written in English and its legal pages were only in German. The German pages remain the ones that apply; each English page says so in its first line, and neither carries a statement the German page does not.

1 October 2026

The study page said the within-cell result is free of engine composition without saying it is not free of period, which only the record page said. It also called the direction of a sensitivity flattering, and printed an outside study’s two figures without saying which is which.

before: “so no engine composition effect can produce it.”; “in the direction that flatters it”; “0.288 and 0.135, against a floor of 0.50 to 0.61”
after: “so no engine composition effect can produce it. It is net of engine and question, not net of period: in this archive each wave is also one sitting.”; “in the direction that raises it”; “0.288 for cosmetic rewording and 0.135 for rewording that adds a constraint, against a floor of 0.50 to 0.61”

A reader of the study page alone met its strongest result without its limit. “Flatters” grades where the sentence only needs a direction. Of the two outside figures only the first is comparable with the paraphrase rule the next study registered, and the registration prints both with their classes.

1 October 2026

The record page said the floor test separates change from same-sitting noise, and that the one company the questions are about is a subject these engines know little about. The first claimed more than the test shows; nothing on the site showed the second.

before: “The floor test separates change from same-sitting noise.”; “a subject these engines know little about and chosen for that reason”
after: “The floor test separates a difference between sittings from same-sitting noise; whether that difference is change over time, it cannot say.”; “chosen because little has been published about it”

In this archive a wave is also one sitting, so the test separates a difference between sittings from the noise inside one; reading that difference as change over time is what it cannot support, and the summary file says so in the next paragraph. The second clause stated what the engines know, which nothing here measured; the reason for the choice is what can be said.

25 September 2026

The method page said recommendation behaviour follows from the other six dimensions, one of which is structured data. Nothing on this site has tested whether the engines it measures read structured data at all.

before: “because it follows from the others”
after: “because it is what the other six are meant to move. For structured data that is an assumption, not a finding: whether the engines this site measures read markup at all has not been tested here.”

“Follows from” states a causal path, and for structured data the path has a step nobody here has measured: whether an engine reads the markup before it answers. Markup that validates, as this site’s own does, is well formed and says nothing about who reads it. The study that would answer it, one capture wave with markup against one without on the same cells, read with the floor test this site already runs, has not been registered. Structured data stays one of the six levers in the sense the page gives the word, a dimension with a concrete fix; what came off is the claim that the fix moves the outcome.

25 September 2026

Three disclosures this site already carried moved to the surfaces that bear them, and a sentence about publication came off the study page. The record page now names the registration and the analysis dates and calls the reading exploratory, and it prints the summary file’s own sentence about sittings beside the claim that the test separates.

before: “exploratory” appeared once on this site, in the last sentence of a paragraph deep in section 5 of the study page; the summary file’s sitting sentence appeared on the study page only, while the record page said “The floor test separates change from same-sitting noise” and stopped there; and the study page read “Nothing here was published before that gap was closed”
after: a seventh limit in the record page’s “What this does not show”, carrying 4 July 2026 and 22 August and the word exploratory, pointing at the method page for both dates; the summary file’s sentence, “In this archive a wave IS a sitting, so a rejection cannot separate drift over elapsed time from answers taken in one sitting resembling each other”, printed on the record page beside the permutation null; and no sentence about publication on the study page at all, whose paragraph now runs from the degrees of freedom straight to “The first analysis plan in this practice is dated 30 August 2026, itself still a draft awaiting ratification”

Two of these three sentences were not wrong. They stood where the reader who needs them does not have to go: the exploratory framing and the sitting limit were on the study page, and the record page carried neither. A disclosure only discloses where it is met. The third was wrong. The gap between a locked question set and an unlocked analysis was never closed for this result; it was succeeded by a draft plan for a later study, so the sentence claimed a virtue for a defect this page still carries. It is cut whole rather than trimmed: the half of it that was true said nothing here had been published, and on a published page that half stops being true first. What it carried now stands on the record page, with two dates a reader can check.

24 September 2026

The study page’s table of prior measurements printed one source’s Jaccard range without saying that the range spans three answer engines.

before: Jaccard 0.29 to 0.50
after: Jaccard 0.29 to 0.50 across three engines

The figure is not wrong and this entry is not a correction of it. Section 5.1 of the cited paper reports platform-level medians one engine at a time, 0.29 to 0.31, 0.33 to 0.40 and 0.50, and 0.29 to 0.50 is the correct envelope of the medians it reports. What the cell did not say is what it had put together: its two endpoints come from different engines and sit 0.21 apart on the same measure. The paragraph under that table argues that a pooled floor cannot tell a stable question on a noisy engine from the reverse, and this cell pools. The number stays. It now says what it is.

22 September 2026

The privacy page dated the acceptance of Cloudflare’s terms with an upper bound read off an e-mail, and the account’s own creation timestamp was available and roughly 59 hours earlier. The note explaining the page’s earlier correction also said no contract had been concluded, which the sentence above it no longer supports.

before: “Die Bedingungen wurden bei Einrichtung des Kontos angenommen, spätestens am 29. März 2026”; “„Spätestens am 29. März 2026“ steht dort, weil eine aufgefundene E-Mail belegt, dass das Konto an diesem Tag bestand, nicht dass es an diesem Tag angelegt wurde”; and “kein Hoster gewählt und kein Vertrag geschlossen war”
after: “Das Konto wurde am 2026-03-26 23:01:52 UTC angelegt, das ist der 27. März 2026, 00:01 MEZ”, with the field it comes from, created_on, and the date it was read named beside it; a limits block whose one remaining limit is which version of the Addendum applied at acceptance, not the date; and “kein gesondert nach Art. 28 DSGVO geschlossener Vertrag vorlag”

A found e-mail proves an account existed that day, not that it was created that day, so the page carried a bound where it could have carried the value. The time and the zone are printed with it because 23:01:52 UTC on the 26th is 00:01 CET on the 27th: a bare date would be either day depending on the reader’s clock. And once the page says the terms were accepted in March, a correction note saying no contract had been concluded contradicts it; what was missing on 2 September was a separately concluded Art. 28 contract and a chosen host, not a contractual relationship at all. The e-mail behind the old bound is not named here, and is not named on the page.

22 September 2026

A section listing four earlier clients came off the record page, and off the Hungarian page that mirrors it. The colophon moved from 10 to 9.

before: “Earlier third-party integration work: idealo, Zalando (gift cards and a gift-card API), ALDI DACH, Deutsche Bank. That work was on the inside of the integrations; the capture tooling above runs against surfaces that were not built to be captured.”, under the heading “9 · Earlier work”
after: no such section on either page, and nothing in its place

Four company names with no dates and no roles is a claim a reader cannot check, on a site whose whole argument is that a claim should be checkable. The same form was removed from a neighbouring section three weeks earlier and the rule was not carried across to this one. It is cut rather than filled in. Nothing takes its place on purpose: a section announcing further studies would be a sentence about work that does not exist yet, which is the other failure this record exists to catch.

17 September 2026

The study page told a reader that the paragraph under a correction mark was unchanged, on the day that paragraph changed.

before: “Corrected 3 September 2026. What follows is unchanged and is the part that mattered: it sits inside the between-wave term…”
after: “Corrected 3 September 2026. What matters is where it sits: inside the between-wave term…”, with the sentence that changed that day carrying its own mark, “Corrected 17 September 2026”

The clause was written when the sentence after it was old. The entry below that ends the study’s upper-bound claim is the change that made it false, the same morning: the finding stopped being called an upper bound and the mark above it was not moved. A correction mark is the one thing on this site a reader cannot check against a file, so it is the one place a stale claim does the most damage.

17 September 2026

Four dated correction marks now name the change each one marks, on the record page and the journey page.

before: one mark on the record page, “Corrected 3 September 2026”, closing a paragraph whose first half had been rewritten on 17 September; and no mark at all on the two sentences the journey page changed that day
after: two marks on the record page, each beside the correction it belongs to; and “Corrected 17 September 2026” on the journey page’s rank-block sentence and on its registration sentence

A mark says when the passage above it last moved, and a reader who remembers an older wording uses it to decide whether to look here at all. Dating a paragraph two weeks before its last correction sends that reader away, and leaving a changed sentence unmarked never sends them at all. The marks were fitted in one pass and the paragraphs were edited in later ones, which is how the dates came apart.

17 September 2026

The record page promised that the study page carries the recipe for every figure it does not mark, and the newest of those figures has no recipe there.

before: “Everything unmarked recomputes from the published rows; the study page gives the recipe for each one.”
after: “Everything unmarked recomputes from the published rows, and the study page writes out the recipe wherever it is more than counting a column. The two subsets in section 1 are not rows below: the study page carries the recipe for the first, and the second is the cells CSV’s within_pairs column at 2 or more.”

52 of 65 was added to the record page that morning and the study page has no recipe for it, so the promise covered a figure it could not deliver. The promise is narrowed to the rows it was written about and the missing rule is given where the figure is printed, which costs a reader one clause rather than a page.

17 September 2026

1,889 came off the record page, together with the caveat added that morning to explain that it is not one of the site’s populations.

before: “it now runs 2,000 permutations over 1,889 cells in under four seconds, no slower than the version it replaced. That 1,889 is a working count from the tool’s own run and is none of the three populations above.”
after: neither sentence

The entry dated 3 September below qualified 1,889 rather than removing it, and that morning’s pass qualified it again. A number that has to be disowned in the sentence after it is a number the paragraph did not need: nothing on this site rests on it, and the paragraph it sat in is about why the analysis code is not published.

17 September 2026

The study page’s table of the headline split by how the floor was measured prints the median difference for the cells whose floor comes from one wave only. That cell was empty. −0.086 is a figure this site has not stated before.

before: an empty cell under “Median difference” in the row “Floor from one wave only”
after: −0.086, the median of delta_text over those 60 cells, read from lib/figures.ts and recomputed from the cells CSV on every build

The table exists to show that the headline is sensitive to how thin the floor is, and the row it was built to expose was the one with no number in it. The value is the worst in the table, against −0.032 on the better-measured half and −0.065 over all 103, so the blank was hiding the end of the range that matters. It is guarded like the figures beside it: the build fails if it stops matching the file, and fails if the cell is typed back into the page.

17 September 2026

The defensible figure stopped being a range. 85 of 103 leads wherever it is printed, and the two better-measured subsets of those 103 stand beside it as subsets. One of them, 52 of the 65 cells whose floor rests on at least two repeat pairs, is new to this site.

before: “the defensible figure is a range, 76.7 to 82.5 percent, and which end applies depends on how much measurement a cell carries”; “82.5 percent sits at the optimistic end of a range whose pessimistic end is 76.7”; “the same way the headline is read as 76.7 to 82.5 percent”
after: 85 of 103 as the headline, with 33 of the 43 cells whose floor is measured from both waves and 52 of the 65 whose floor rests on at least two repeat pairs printed beside it rather than bracketing it; and 76.7 named as the share on a subset, carrying its own uncertainty and not the end of anything

A subgroup estimate is not a bound. 33 of 43 is the same count over a smaller and better-measured set, so it carries its own uncertainty rather than marking where the truth stops, and printing 76.7 to 82.5 as a range told a reader the answer lies somewhere between them, which the design does not support. 52 of 65 was recomputed from the cells CSV for this change rather than carried across: of the 103 cells with a floor, 38 rest on a single repeat pair, which leaves 65, and 52 of those moved. It is checked against the file before every build, as 85 of 103 and 33 of 43 already were.

17 September 2026

The study stopped claiming that the floor test settles the confound it names. The floor is measured inside one sitting and the effect across sittings weeks apart, and the summary file says in its own words that a rejection cannot separate the two.

before: “The floor test fixes that in one move.”; and “That is a within-cell comparison, so it is the one number here that says anything about elapsed time.”
after: “The floor test separates change from same-sitting noise.”; “net of engine and question. It is not net of period.”; and the summary file’s own sentence printed beside the permutation result, “In this archive a wave IS a sitting, so a rejection cannot separate drift over elapsed time from answers taken in one sitting resembling each other.”

The sentence was in the published summary file the whole time and the pages said the opposite of it, which is the case the pages’ own rule covers: where the file and the page disagree about anything the file covers, the file is right. Elapsed time is also confounded with July against August in this archive, so a within-cell comparison is net of engine and question and of nothing else. No figure moved. What moved is what the figures are allowed to mean.

17 September 2026

The study stopped calling its finding an upper bound on what the engines did.

before: “every mechanism of that kind pushes the result in the direction this page reports, so the finding is an upper bound on what the engines did”
after: “the unrecorded between-wave factors named here bias the reading upward. That is a direction, not a bound: sampling error can still sit on either side of it.”

A bound says the true value is at most this one. What the unrecorded between-wave factors support is a direction: they push the reading up, and sampling error still sits on either side of the pushed value, so the reported share is not a ceiling. The phrase also stands in the reasoning of the account-state entry below, dated 3 September, where it is now wrong for the same reason; this entry is its correction.

17 September 2026

The table of shares by the gap between waves stopped being called a dose-response.

before: “What the spread does give is something better than the headline: a dose-response.”, over a table captioned “Share moved, by the gap between the two waves compared.”
after: “a breakdown by gap”, with the summary file’s own warning beside it, “Bands below compare DIFFERENT cells and different engines to each other, not the same cell over time”, and the caption ending “Between cells, not the same cell over time.”

A dose-response is a claim about one thing responding to more of another. These bands compare different cells and different engines to each other, so a share that rises across them can be produced by which cells fall in which band. The published summary file names that trap in a field of its own, and the page published to describe that file called the table the opposite of what the file says it is.

17 September 2026

The record page said the tie-break spread is wider than the confidence interval printed beside it. It is the narrower of the two.

before: “the spread is wider than the interval printed beside it below”, and “the tie-break spread is wider than the interval”
after: “reaches past one end of the interval”, in both places

The spread runs −0.042 to −0.066 and the interval −0.058 to −0.031. What is true is that one end of the spread lies outside the interval: −0.066 sits below −0.058, while −0.042 sits inside it. So the sentence pointed at something real and described it as the opposite of what it is, once in the first paragraph of the record page and once beside the figure. All four values are in the published files and are checked against them before every build; the comparison between them was not.

17 September 2026

Half of a sentence about the quartile alternatives was wrong, on the study page and on this one.

before: “two other conventions give 88.5 and 100.0, or 84.0 and 92.3, for the upper two bands”
after: “88.5 and 100.0 for the upper two bands, or 84.0 and 92.3 for the middle two”

Under the convention that produces them, 84.0 and 92.3 are the shares of the second and third blocks and not of the third and fourth. The other pair does belong to the upper two, so the sentence was split rather than rewritten: the true half stands and the false half is corrected. Both pages carried the sentence and both are corrected in the same series, because correcting one copy and not the other is how this class of error survives a check.

17 September 2026

The seven capture waves stopped being called two clusters. The body of the study page was corrected first, and its headline and the machine-readable copy of that headline went on carrying the old reading.

before: “they fall in two clusters”; “Seven waves in two clusters is a July-to-August comparison, not a trend”; and “more between two capture clusters than they do when asked twice in a row” in the headline and in the structured data
after: “they fall at three points, early July, mid-August and late August” in the body, with “Seven waves at three points” beside it; and a headline, with its structured-data copy, that makes no claim about the structure at all: “more between capture waves than they do when asked twice in a row”, the words the record page already uses

The dated wave table on the study page is what decides it. Grouped at the two largest gaps between consecutive waves, the seven make three points, and the widest gap inside any of the three is smaller than the smallest gap between them. Grouped at the largest gap alone, they make two clusters, and the August one is then wider inside itself than the gap that is supposed to separate it. Two clusters was the older reading and it survived in the headline and in the structured data after the body had been corrected, which would have left the page contradicting its own title on the surface a search engine reads. The headline does not gain the corrected structure; it loses the claim: three points is a fact about the wave table and it belongs beside the table, and a title that carries it has to be re-edited every time the archive grows.

17 September 2026

The capture window is stated as what the dates actually are, on all three pages that print it.

before: “seven capture waves between 4 July and 25 August 2026”
after: “seven capture waves whose first captures fall between 4 July and 25 August 2026”

Three of the seven waves ran across several days and the last ran on into 27 August, which the study’s limits already said. The date printed for each wave is its first capture, so the old form described the table rather than the archive. The qualified wording was already on the study page and is now on the record and journey pages as well.

17 September 2026

The study page named the wrong wave as the one that contributes no cell.

before: “Only six of the seven contribute a cell: the fourth, makefield-cycle0-20260705”
after: “the second”

The wave table three lines below the sentence lists makefield-cycle0-20260705 second of the seven. The reason it contributes nothing was right: one draw per cell and a single draw makes no pair; the ordinal was wrong, directly above the table that contradicts it.

17 September 2026

The download card described the summary file by a count of sections it no longer has.

before: contents: “9 sections”
after: contents: “24 top-level keys”

The published summary file has 24 keys at its top level. Nine described an earlier shape of the file, and nothing moved when the file grew, which is what a hand-typed count with no check does. The card’s count is now compared against the published file before every build, so the two cannot drift apart again without the build saying so.

17 September 2026

The record page’s limitation about scoring was aimed at something this site does not do.

before: “No second scorer has rated anything, so every score here is one person’s reading.”
after: “Every number here comes from one pipeline, written by the person reporting it and run over an archive captured by the same person.”

Nothing here is scored by a human rater, so a second scorer was never the missing control. The sentence implied a rating step that does not exist and left the real dependency unnamed. What every figure on this site depends on is one pipeline, written by the person who reports its output, over an archive that person captured. That is the wider claim and it is the one a reader can weigh against the published files.

17 September 2026

The method page said the raw read is recorded by hand, while the record page says the captures come from a Chrome extension. Only one of those can be true of the same step.

before: “The raw read is recorded by hand: the answer text, the sources, a screenshot, and the conditions the answer was produced under.”
after: “What is recorded of the raw read is” the same four things, followed by “The capture itself is not manual: a Chrome extension I wrote performs it and writes the rows, while a person runs the session and, as below, does not score while it is running. Where the line between the two falls, step by step, is in no published file, so a reader has this page’s word for it and not a record.”

Two pages of this site described one step in incompatible ways, and the page a reader is sent to for method said hand. The extension writes the rows and a person runs the session. Where exactly the line between them falls is written down in no file anyone can read, so the correction states what the two pages jointly support and names the rest as unrecorded rather than filling it in from memory.

17 September 2026

The journey page said the next study is registered. Its registration is a draft awaiting ratification, which the study page and the method page already said and the journey page did not.

before: “The study those rules were written for is registered and has not been captured.”
after: the same sentence with the registration named as a draft awaiting ratification, in the form the other two pages use

The registration file dated 30 August 2026 carries a status of draft, says it is not locked until it is ratified and the ratification is dated, and qualifies its own commitments as standing before ratification. Registered, said without that, claims a commitment that has not been made yet. The study page and the method page both carry the fuller form, so the journey page was the thin one, and a reader comparing the three would have found the strongest version of the claim on the page carrying the least detail.

17 September 2026

The record page’s list of four other repositories, with a commit count for each, came off the site.

before: a section headed “Other repositories” naming four projects and giving each a commit count, none of them linked to anything
after: the section deleted, and the two sections that followed it renumbered

The counts could not be checked. Nothing on the page or off it opens any of the four, so a reader could neither read the history nor recount it, on a site whose argument is that a claim arrives with the thing that checks it. I decided on 10 September that they come off rather than gain links, because linking them puts working code on a page about measurement and that is a different decision from this one. Nothing else on the page depended on the section.

17 September 2026

A promise that the journey page names two places where the instrument caught its own author was withdrawn, because the page names none.

before: “Seven stations from the first question to the current claim, including the two places the instrument caught its own author.”, in the journey page’s visible text, in its route description, on the error page, and in the hand-written structured data
after: the sentence ending at “the current claim”, in all four

No mark on that page says which two stations they are, and the line nearest the promise says the opposite of it. Paying the promise means choosing which two, which is an editorial decision and not a repair, so it was withdrawn instead. It stood in four places for one sentence, two of them inside the same route, because that route’s description is written once for the browser and once for the structured data. A sentence duplicated that way is corrected in every copy or in none.

17 September 2026

The colophon said this site is set in one typeface.

before: “Set in one typeface, no analytics”
after: “Set in a system serif and a system monospace, no analytics”

The body is a serif and the labels, values and code are a monospace, which is visible on the page the sentence sits on. It is a small claim and it is the kind this site is least entitled to get wrong: a statement about itself that a reader checks by looking up.

17 September 2026

The colophon offered the code that produced the measurement. The code is not published, and the same page says so twice.

before: “what the measurement cannot tell you, and the code that produced it”
after: “and the files it produced”

Two paragraphs of the same page read “It is not published” and “The tool is not public”, so the colophon promised a reader something its own page withholds. What is published is the data the run produced, and that is what the line names now. The reason the tool is withheld is unchanged and is on the same page.

17 September 2026

The first paragraph of the site leads with the finding rather than with the verifier failure, and the page carries an address to write to.

before: “My verifier failed the resulting archive on four of its eleven checks when it ran on 29 August 2026, a read that has not been reproduced since, and the figure below is the second one I arrived at”, as the second sentence on the site
after: the finding and its consequence in the first paragraph with the eight engines named, the verifier result left in section 6 where it already stood, and a contact line

The verifier result is not withdrawn and not softened. It is in section 6, with its date and the note that it has not been reproduced, exactly as it was. What changed is that it is no longer the second sentence a reader meets. An earlier entry below records the same sentence being corrected in the same position on 3 September, for its missing qualifier: the opening paragraph is the most-read one here and this is the second time it has had to be edited for what it puts first.

17 September 2026

Two clauses left over from earlier corrections came out: the last page still reporting shares by quartile of the floor, and a comparison to a six-wave version of the study that no page explains.

before: “Every share is now reported by quartile of the floor”; and “The waves fall at three points, early July, mid-August and late August, which allows one test the six-wave version could not run.”
after: “by rank block of the floor”; and the sentence ending at “late August”

The bands are cut by rank and not at the column’s quartile values, which is the correction of 1 and 3 September; the journey page was the last surface still using the old word for them. The six-wave version is an earlier scope of the same study and is named nowhere a reader can reach, so the clause claimed an advantage over something invisible. One use of the old word stays on purpose: the journey page calls the finding itself “the quartile split”, which is the name it was found under and the name this record lists it by.

17 September 2026

Three pointers sent a reader to the wrong place. No stated figure changed in any of them, which is why they are one entry and not three.

before: “the first file in the download above”, under a figure the download sits below; “the same direction and the low end of the range below”, where that range is above; and “The study page names which is which, one by one”, printed on the study page itself
after: “above” and “below” swapped to match the layout, and the study page’s copy of the download block reading “This page names which is which”, with the record page’s pointer to the study page unchanged

A pointer is not a claim about the world and none of the three moved a number, so they are listed together. They are listed at all because a reader who followed one and found nothing has a right to know it was the page and not their reading. The third is the one worth naming: the download block renders on two routes, so on the study page it sent a reader from that page to that page. It is fixed by having the block say which page it is on, rather than by deleting the sentence, which would have cost the record page a pointer that works.

17 September 2026

This page stated two things about the site that were not true of it: that the record page marks five rows as read out of the summary, and that every tie in the gap-pairs file is exactly two rows, which the entry directly above that one had already retracted.

before: “five rows marked ‘from the summary’”; and “Every tie in the gap-pairs file is exactly two rows”, still asserted as fact inside an entry sitting one place below its own retraction
after: “four rows marked ‘from the summary’”; and that sentence marked false where it stands, with the 119 tied groups and the 20 holding three rows named beside it

The record page carries four rows marked from the summary and says four in the sentence above them, so this page was the only surface still saying five and nothing had corrected it. The tie universal is worse: it was retracted on this page, on the same day, by the entry printed immediately above the one that goes on asserting it, so a reader meeting them in order was told a thing and then told it was false without either entry acknowledging the other. A record that contradicts itself inside one screen is the failure this page exists to catch, committed by the page. Both were found by a review of this site on 10 September 2026, and neither is a figure on any other page. The stale line saying one placeholder is left on this site went the same way: the placeholder was filled on 10 September and that has its own entry above.

10 September 2026

The privacy notice now names the contract with the host and the basis for transferring data to the US. Until this date that line was a visible placeholder.

before: “[VOR DEM LIVEGANG EINTRAGEN: die tatsächliche Grundlage der Übermittlung, wie sie im abgeschlossenen Auftragsverarbeitungsvertrag steht, und das Datum des Abschlusses]”
after: Cloudflare’s Data Processing Addendum, which Cloudflare incorporates by reference into its Self-Serve Subscription Agreement, accepted when the account was set up and no later than 29 March 2026, current version 6.4 of 3 April 2026; transfers to the US on the EU-U.S. Data Privacy Framework and the Commission’s adequacy decision of 10 July 2023, with the Commission’s standard contractual clauses under section 6.2 of the addendum if the certification lapses; the certification checked on 10 September 2026 in the US Department of Commerce list, status Active

The line was a placeholder on purpose, so that no transfer basis would be named before anyone had read which one the contract actually carries. It was filled once the addendum and the terms had been read in full and the certification had been looked up in the official list. “No later than 29 March 2026” goes exactly as far as the evidence: an e-mail shows the account existed that day, not that it was created that day.

3 September 2026

The Impressum has a real postal address, and it is a home address. It also stopped explaining Impressum law to its reader and stopped asserting an exemption it could not meet.

before: “[STRASSE UND HAUSNUMMER]”, “[PLZ] Berlin”, and a note reading “Eine E-Mail-Adresse ist Pflicht. Eine Telefonnummer ist es nicht, solange ein zweiter schneller Kontaktweg besteht”; the Datenschutz controller section carried two placeholders of its own
after: Britzkestraße 9, 12347 Berlin, on both pages, with the contact address filled; and one line saying how to reach me rather than an argument about what the law requires

My decision, taken against a legal review that rated the irreversibility of publishing a home address at its highest severity, and taken in preference to a service address. The reason is the one that review could not settle: a c/o address counts only where the provider holds a Zustellungsvollmacht under § 171 ZPO, and no court has settled that this satisfies the Impressum address, so a mailbox company buys an unsettled question rather than an answer. The exemption sentence went because a review found the page claiming a second fast contact route and then providing none; whether a telephone number is therefore required is with the Fachanwalt and is not answered by removing the claim. One placeholder was left on this site that day, the Art. 28 line; it was filled on 10 September 2026 and has the entry above.

3 September 2026

The 108 cells behind the strongest figure on this site were never the 103, and are not a subset of the 111 either. Two of them span the question change that twelve cells are excluded for.

before: “measured inside cells so that no engine or question composition effect can produce it”, with 108 printed on three routes and its population stated on none
after: “measured inside cells, so no engine composition effect can produce it”, with the 108 traced to the 123 cells the gap-pairs file covers, 100 of them inside the 111 and 8 outside, and the two straddling cells named and costed

Recomputed from the gap-pairs file rather than read out of the summary, which is the whole reason that file was published on 3 September. It covers 123 cells, because a cell the floor test drops for a changed question can still hold wave pairs whose question text was identical, and only those pairs are in the file. Fifteen carry one distinct gap and fall out, leaving 108. Restricted to the 100 that are also in the floor test, the median is −0.042 with 79 of 100 negative. Two cells are worse than mismatched: B3 and B4 on ChatGPT take their shortest gap from the question as first written and their longest from the corrected one, so the statistic subtracts one question from another and “same question, same engine” is untrue of them. Dropping both gives −0.046 over 106, with 85 negative and 21 positive. The published figure survives all three readings and stays inside the tie-break range already printed beside it. The sentence ruling out a question composition effect did not survive, and it was the clause doing the persuading.

3 September 2026

The study’s quartile table printed a Range column that does not reproduce its own shares, and the recipe retracted on 1 September was still live in two other places, one of which renders on two routes.

before: a Range column reading “up to 0.200 / 0.200 to 0.266 / 0.266 to 0.321 / above 0.321”; and “split at the quartiles of the floor itself” on the journey page and under the figure
after: a column headed “Floors it contains” carrying the observed span of each rank block, 0.0357 to 0.1985 and so on; and “sorted by their own floor and split into four equal rank blocks”

Binning by those printed boundary values gives 57.7, 81.5, 95.8 and 96.2. The dataset README calls that the trap, and this table was the thing setting it: a reader who did the arithmetic correctly from the column beside the shares would have got different numbers and concluded the page was wrong. The four shares themselves are unchanged and recompute from the CSV under the rank-block rule. The retracted recipe surviving in the figure caption mattered more than its length suggests, because that component renders on the record page and on the study page, so the withdrawn wording was reaching a reader twice for each time the correction did. The study page also sent a reader to section 3 for three values printed in section 2.

3 September 2026

The record page printed two bootstrapped intervals and a permutation null in the same list and the same face as figures anyone can recompute. Its engine list did not say which population it counts, and the verifier table showed 26 twice with no word on whether that is one number.

before: one undifferentiated counts list; an engine list with no population; two rows reading 26
after: four rows marked “from the summary”; the engine list stated as all 111 comparable cells and pointed at the study’s list over the 103; and a line saying the two 26s count different objects and cannot be reconciled from anything published

The split between what recomputes and what is read out of the summary was already stated on the study page, under the download block and on this page, which is three routes away from the page most people read, and a reader forms the impression from the list in front of them. The engine list sums to 111 and the study’s sums to 103, so the two looked like they disagreed when they are counting different things. On the two 26s: cross-run references counts pictures recorded under one run and referenced from another, unkeyed observations counts answer rows with no cell address, the verifier output does not say whether the sets overlap, and the archive it read is not published, so the answer is that it cannot be settled from here rather than that it is a coincidence.

3 September 2026

The journey page said six rules were each paid for by an error it names, where two are. It promised that unmarked text is never an unlabelled fact while its densest station carried five facts and no marks. And one observed mark reported a dead build in the present tense and closed on an inference.

before: “Six rules, each one paid for by a specific error above”; “Unmarked text is ordinary prose, never an unlabelled fact”; “A verifier written against the archive fails it on four of eleven checks … The instrument had recorded that it had not followed its own protocol.”
after: two rules attributed to the errors that bought them and the other four disclaimed; a marks rule narrowed to what the page does, with station 4’s two figure paragraphs marked observed; and the verifier mark carrying its date and build id, with the inference moved into an interpretation mark

Rules 2 and 5 answer the registration gap and the field read as what it implies, both named in station 4. Rules 1, 3, 4 and 6 have no counterpart anywhere on the page, so six rules were wearing the authority of two. The marks rule was falsified by the page stating it, which is the same failure as a claim without its check. And the verifier line said fails, present tense, about build 2136073e, which the record page says is not the build in the archive today and whose counts have not been reproduced since: a mark reserved for what can be read off a named file was carrying an undated claim and ending on a conclusion.

3 September 2026

Three sentences on the method page took back an admission in the same breath as making it. One claim about what this site offers was wider than the change record allows, and the registration dates were called public when three of eleven are.

before: “which is a stage and not a defect”; “so no harm was done”; “it is declared here rather than discovered by a reader”; “being checkable twice is the only thing this site is offering”; “The dates of the registrations are public”
after: the three admissions without their softeners; the offer narrowed to the published rows, with the change record named as a record and not a check; and three of the eleven dates stated as the three that are public, with one file carrying no date at all

A hedge attached to an admission takes back what the admission gave, and all three sat on the page whose subject is where the discipline around this work failed. Being checkable twice is what this page is for, and this page is written by the person who makes the changes, which it says about itself two blocks up; the part a reader can check without the author is the published rows. And section 4 of the method page publishes three dates and says one file carries none, so a sentence two sections later contradicted the page it sits on.

3 September 2026

A parenthetical describing two other companies’ arrangements, an unfalsifiable line in the colophon, a fourth denominator with no explanation, the verifier result stated without its qualifier in the first paragraph on the site, and an error page that asked for an identifier and named nowhere to send it.

before: “ALDI DACH (a partnership with Google and Apple)”; “Written by hand”; “2000 permutations over 1889 cells”; “My verifier fails the resulting archive on four of its eleven checks”; and an error page with a digest and no address
after: “ALDI DACH”; a colophon without the claim; 1,889 named as a working count from the tool’s own run and none of the three populations on the page; the verifier result carrying its date and the note that it has not been reproduced; and an address on the error page

The parenthetical put two companies’ commercial arrangements into a record of one person’s work, which is the failure the namesake removal corrected earlier the same day. Written by hand cannot be checked, in a colophon whose other claims can. The 1,889 read as a fourth population beside 111, 103 and 108 and is none of them. The opening paragraph is the most-read sentence here and was the one place of five where the verifier result appeared without the qualifier the other four carry. And an identifier a reader cannot send anywhere is a request that cannot be acted on, the same shape as the instruction to read unpublished files that came out that morning.

3 September 2026

Account state was described as a control on three pages. It was never recorded, so it was never a control. And the study’s headline said the engines “changed”, where what is measured is wording overlap.

before: “account state is therefore held constant and not known”; and “Answer engines changed more between two capture clusters than they change when asked twice in a row”
after: “neither controlled nor observed … nothing held it, and it may have varied between waves”; and “Answer engines reworded their answers more between two capture clusters than they do”

Held constant and not known are two incompatible claims in one sentence: the first makes account state a control, the second makes it a confound. This is the same class of error as the sign-in correction of 31 August, which is why it survived that correction: the inference was replaced and the word around it was not. The consequence was already stated correctly and is unchanged, that it sits in the between-wave term and inflates the reading rather than damping it, which is what makes the finding an upper bound. The headline is the sentence most likely to be quoted, and Jaccard over content tokens sees wording overlap and not agreement, which the study says four paragraphs below the headline that contradicted it.

3 September 2026

The reason given for withholding the capture tool was replaced, and the repository-history section was rewritten around what a reader can actually do with it.

before: a reason that stated an unchecked doubt as though it were a finding; and a section listing five internal filenames with the instruction “read those instead of the commit log”
after: the tool is withheld for the same reason the question set and the scoring anchors are; and a section saying plainly that the instrument’s history is gone, so there is no way to show from here that the tool on 4 July was the tool on 25 August

A review ruled the old reason a rationalisation, and said its own earlier passes had not examined it: it stated an unchecked doubt as though it were a finding, which is the thing this site exists to catch. The repository section had the same shape: it told a reader to read five documents that are not published, from a repository that is not published, which is an instruction nobody can follow. What survives is the limitation itself, which is real and now stated as one.

3 September 2026

A third party’s domain was removed from the study page, and the reason twelve cells are excluded was restated as what it actually was: an instrument error, not a rename.

before: a sentence naming the other company’s domain and saying it “became” mine in two of the eighteen questions (the domain is not reprinted here, for the same reason it was removed)
after: two of the eighteen questions had been written against the wrong domain, a namesake business with no connection to this one; the namesake is not named

The old sentence read as though a domain of mine had been renamed. It was not. Two questions named a business I have nothing to do with, and the first wave asked eight engines about that business by mistake. The error was caught on the day of that wave, 4 July 2026, and both questions were corrected. Naming the namesake put a real company into a research narrative it had no part in and gained nothing from, which is the kind of thing this site is supposed to catch rather than commit. The published data is unaffected: the only rows for those two questions compare two later waves, both of which used the correct domain.

3 September 2026

The reason the analysis tools are held was corrected. It named the wrong reason twice in two days.

before: “clean of client material”, then “clean of material from earlier client work”
after: an export that can be cut to this run alone, because the archive also holds runs over other companies’ brands, captured without a brief from them

Checked against the rule that actually governs it rather than against memory. Of the 38 runs in the archive, six are over other companies’ brands with no brief from them and nine are held behind a consent gate; the rest are self-audit, instrument and internal research. So the blocker was never a claim about clients, past or present. The first wording asserted current clients, which a review flagged. The second asserted past client work as the blocker, which the governing file does not support. This one says what is in the archive and why none of it can go out.

3 September 2026

A third pass over the same paragraph found that the correction written earlier that day introduced a false statement and a sentence that had lost its subject.

before: “Every tie in this file is exactly two rows”, and a clause reading “The point estimate does not” sitting directly after a sentence saying the point estimate is unchanged
after: “Every tie the rule actually touches is exactly two rows”, with the 119 tied groups and the 20 holding three rows stated; the orphaned clause moved back to the sentence it belongs to

The universal was false. The file holds 119 tied gap-day groups within cells and 20 of them hold three rows; the claim is true only of the 37 groups at a cell’s shortest or longest gap, which is the set the rule reaches. It was false inside the paragraph whose subject is stating rules precisely, in the file this site declares authoritative over its own pages. The orphaned clause was collateral from moving a sentence during the earlier edit. Three passes each found one narrow thing on this one figure, and the third found only defects the second pass’s fix had created, so the paragraph is now finished rather than improved further.

3 September 2026

The paragraph that publishes the pairing rule named a fifth alternative that is not a rule, and the record page showed the figure with a tighter interval than the site’s own analysis supports. Both were corrected on the day they were written.

before: “the median of the tied rows, -0.0658 and 93”; and the record page listing −0.047 with a 95 percent interval and no sensitivity range
after: “the mean of the two tied values, -0.0582 and 90”, with the reason printed; and the record page carrying −0.042 to −0.066 and 83 to 93 of 108

Every tie in the gap-pairs file is exactly two rows, so a median over them has no middle row. ⚠ That first sentence is itself false and the entry above corrects it: the file holds 119 tied gap-day groups and 20 of them hold three rows, so the universal is true only of the 37 groups the rule reaches. Read one way, the median repeats the alternative named one clause earlier; read the other, it gives a figure that was not printed. Naming an under-specified rule inside the paragraph that exists to insist rules must be specified is the quartile-band failure of 1 September one level down. And the record page and the study page disagreed about how firm one number is, which the change record had already corrected once that week somewhere else. The range endpoints and the published point estimate are unchanged by both.

3 September 2026

A third file was published, and it moved the site’s strongest figure from stated to checkable.

before: two files; the within-cell gap test could not be recomputed and its pairing rule was written down nowhere released
after: three files; floor-test-gap-pairs-08284347.csv, 957 rows, one per wave pair, with the pairing rule and its tie-break in the README

A review found this to be an over-withholding failure rather than an over-sharing one: what blocked the check was a rule, not an instrument. Publishing it also measured the ambiguity it warned about. A cell can hold two wave pairs at the same gap in days and 37 of the 108 do, so five reasonable tie-breaks give medians from -0.042 to -0.066 and counts from 83 to 93 of 108. The direction holds under all five; the point estimate does not, and the published value is the second most conservative of them.

3 September 2026

A review ran over this site and the dataset, and three of its findings changed published strings. The attribution line no longer names a company, two pages no longer assert current clients, and the summary file gained a sixth caveat, which moved its checksum.

before: attribution: “Denes Csaszar, Makefield”; “held until a single-scope export comes out clean of client material”; summary sha256 aca27635…
after: attribution: “Denes Csaszar”; “clean of material from earlier client work”; summary sha256 cf21d8a9…, 11,684 bytes

The gate reads the site in its current phase, where the author publishes as a person and not as a business. An attribution line is where the actor is named, and a present-tense sentence about cleaning client material asserts that clients exist now. The caveat was added because the summary file travels alone under its licence, and until today it carried a term-presence count of zero for the author’s own domain without the sentence that says the subject was chosen to be unknown by design. A reader holding the file alone got the zero without the reason for it.

3 September 2026

The study page’s list of what is read out of the summary rather than checked against it went from six families to four. Publishing the third file is what moved two of them, and the correction was written into the study page and listed nowhere.

before: six families, including the within-cell gap test and the count of seven waves
after: four: the permutation null, the twelve-cell coverage check, the two bootstrapped intervals, and the count of twelve excluded cells

Section 3 of this page carries the same correction for this page’s own list and has since 3 September; the study page’s version of it was retold on the study page and recorded nowhere else, so a reader who met the list there and came here to find out when it changed found nothing. The two that moved did not move by argument: the gap-pairs CSV published that day makes the within-cell gap test recomputable, and the seven wave labels are readable off the same file. This entry is written on 17 September 2026, two weeks late, because a review of the site found the inline note and this list did not have it.

2 September 2026

The method page compared the twelve-cell deviation against the confidence interval and stated the factor against the interval’s half-width instead of the interval. The correction was written into the method page and listed nowhere.

before: the two differ by 0.0007, about forty times smaller than the confidence interval
after: the interval on that median is 0.062 wide, about ninety times larger

Forty is the factor against the half-width, 0.031. The interval as this site prints it on every other surface is −0.092 to −0.030, which is 0.062 wide, and 0.062 over 0.0007 is about ninety. So a reader dividing the two numbers printed beside the sentence would have got a different answer from the page, which is the quartile-band failure in a second place, and the correction note on that page says so. Nothing is withdrawn by it: the difference is far too small for this data to resolve at either factor. This entry is written on 17 September 2026, two weeks late, for the same reason as the entry above it.

2 September 2026

The carve-out under the download block was completed. It named one exception and there are several.

before: the verifier’s counts are read off the archive, which is not published
after: the CSV can be recomputed against; the permutation null, the within-cell gap test, the twelve-cell coverage check, the two bootstrapped intervals, the wave count and the excluded-cell count are read out of the summary file rather than derived; the verifier’s counts are in neither file

The narrowing on 1 September fixed the word ‘every’ and left the list short. The study page had already named three measurements that do not recompute, and the block under it named only the verifier. Two more came out of checking it again: the CSV yields six wave labels where the summary lists seven, and the twelve excluded cells are stated in the summary and absent from the CSV, which holds 111 rows of which 8 have no floor.

1 September 2026

The request counter on this page was dated, and the offer beside it was narrowed.

before: Ask and you get the verifier and the hashes it checks against. Requests so far: 0.
after: the same offer, plus what it does not buy, plus a date on the count

The count is typed by hand, so it is the one number on this site that no file backs. Left undated, it drifts in the flattering direction: a stale zero reads as nobody having asked. And the offer, read alone, sounded like it closed the gap the paragraph above says it does not.

1 September 2026

The claim that every figure on this site can be recomputed from the two published files was narrowed, in two places.

before: every figure on this site / on this page can be recomputed from these two files
after: the measurements can; the verifier’s counts are read off the archive, which is not published

Checked rather than assumed. The per-engine cell counts, the 85 of 103, the 33 of 43 and the median difference all reproduce from the CSV. The verifier’s five figures are in neither file. One of the two overclaiming sentences had been live since before the change record started.

1 September 2026

The study page said four of its figures recompute from the CSV. Five do, and two of the recipes it gave were too loose to follow.

before: four figures, with the gap table described as gap_days banded at 14 and 40 and the quartile table as the column split at its own quartiles
after: five, each written out in full, with the population stated as the 103 cells that have a floor

All fourteen numbers were recomputed from the CSV and all fourteen match, so the claim was true and too modest. But a reader following the gap-table recipe over all 111 rows gets 28 of 40 and 49 of 59 instead of 28 of 35 and 49 of 56, and the quartile recipe did not determine an answer at all.

1 September 2026

The dataset README now states how the four quartile bands are cut.

before: shares are computed at the column’s own quartiles
after: equal rank blocks of 26, 26, 25 and 26 over the cells sorted by noise_sim_text

The shares reproduce from the CSV, but four plausible binning rules give four different answers for the upper two bands, and the file named none of them. Binning by the three boundary values the pages print does not reproduce the shares at all.

1 September 2026

A revision date was added to the method, study and journey pages.

Only the record page carried one. Three of the four public pages stated no date at all, so a reader of a page full of figures could not tell when the figures were last touched without going to the repository.

1 September 2026

The record page’s structured data claimed a different revision date from the page itself.

before: dateModified 31 August 2026, while the colophon and the sitemap said 1 September
after: one date per route, in one file, read by all three surfaces

The date lived in three places and two of them were updated. Google’s sitemap documentation uses a lastmod value only when it is consistently and verifiably accurate, so three dates that disagree are worse than none.

1 September 2026

The journey page was added, and the record page links to it from its opening paragraph.

The record states a figure and the study states how it was computed, and neither says whether it came out that way the first time. It did not.

31 August 2026

Every claim about whether the capture sessions were signed in was removed from all three pages and from the dataset README.

before: the sessions were signed in throughout, inferred from the capture mode
after: the capture mode was fixed per engine and unchanged across every wave, so it was controlled; sign-in state was not recorded and cannot be inferred from the mode

The inference was checked against the live surface and failed: chatgpt.com/temporary is reachable while signed out. The field that looks like it records sign-in state is an alias for one of five capture modes. An inference is not a record.

31 August 2026

The quartile shares in the dataset README were restated at one decimal.

before: 56, 84, 92 and 96 percent
after: 57.7, 84.6, 92.0 and 96.2 percent

The rounded figures were printed in one place and the exact ones in another, so the same four numbers appeared twice on this site in two forms.

3 · How to check any of it

The entries above that changed a number name both values, so any of them can be checked against what the pages say now. Beyond that, the answer depends on which figure.

Recomputable from the two CSVs: the 103 cells and the 85 that moved, the per-engine cell counts, the 18 questions and the 111 rows, the noise floor and its quartiles, the between-wave median, the median difference, the wave dates, the gap table, the margin counts, the quartile shares and the split between cells whose floor comes from one wave and from both. Checked on 1 September 2026, and again on 2 September, by recomputing each from the cells CSV rather than by trusting the summary. And from 3 September the within-cell gap test as well, from the gap-pairs CSV published that day: median −0.0465 over 108 cells, 87 negative and 21 positive, reproduced exactly under the pairing rule now written into the README.

Stated in the summary file and not derivable from any of them: the permutation null, which needs the similarity matrix; the twelve-cell coverage check, which needs the registered draw targets; the two bootstrapped intervals, which carry their seed and their replicate count but not the resampling code; and the count of twelve excluded cells, which appears in neither CSV. Reading a figure out of the summary is not checking it.

⚠ This paragraph has been corrected twice in two days, and both corrections are worth seeing. On 2 September it replaced a list that called the permutation null recomputable, which the study page had already contradicted: two pages of this site disagreed about what was checkable. On 3 September it went from six items to four, and not by argument: the within-cell gap test and the count of seven waves both recompute from the file published that day.

Not in those files: everything the verifier reports. The counts of ambiguous picture references, duplicate frames and cross-run references, the re-hash totals and the build id are read off the archive, and the archive is not published. Those figures are documented, not independently recomputable. The verifier’s source is not public; it can be sent on request. The archive it reads is not published at all. So a reader can learn what is counted, by asking, and still cannot recount it, and that stays true until either the archive or a signed verifier output is published. Neither is today.

Requests go to archive@denescsaszar.com, and what you get is the verifier’s source and the hashes it checks against: enough to see how the counting is defined, not enough to repeat it. Requests as of 1 September 2026: 0. That count is kept by hand, so it goes stale without anything failing. The date is there so a reader can see how stale.

One figure that is recomputable and still was not reproducible

The quartile shares reproduce from the CSV exactly, but only under one of several plausible ways to cut a column into quarters. Two rules give the published 57.7, 84.6, 92.0 and 96.2; two others give 88.5 and 100.0 for the upper two bands, or 84.0 and 92.3 for the middle two. The files did not say which rule, so a reader doing the arithmetic correctly could land on different numbers and reasonably conclude the page was wrong. The rule is now written into the dataset README.

⚠ Corrected 2 September 2026. The second alternative read 81.5 and 95.8, which is the result of binning by the printed boundary values, and that case is the separate trap described next. So one convention was counted twice and another was left out, and this page and the study page gave different numbers for the same fact. The study page was right.

Worse and more specific: binning by the three boundary values the pages themselves print, 0.200, 0.266 and 0.321, does not reproduce the shares. Those are the column’s quartile values, not the cut points that produced the bands.

The figures stand on the record and in the study; this page says only what moved.

Last revised 31 October 2026.