Every number, sourced

2026-09-07 · 3 corrections

• methodRun 1 of the swap audit (1 September 2026) recorded as claude-opus-5 and run 1R (4 September 2026) as claude-fable-5, the pair described as two model families and the replication as a family-generalisation check.

  Now: Both runs were claude-fable-5. The six reader sessions of run 1 each carry the model field claude-fable-5, and all eighteen reply strings held in result-run1.json appear verbatim in those six sessions, one reader per session. What replicated on 4 September was the same model three days later, with replies collected as free text instead of three structured fields. Corrected in version 1.0.1 of the record and in the two research files, with the value first recorded kept beside the corrected one, dated.

  Our own error. The working session had been switched to claude-fable-5 about two hours before run 1 was launched, and the assistant wrote its previous model name into the result file; nothing checked that field against the reader sessions before deposit. Found by the project's own audit of the record on 5 September 2026 and verified against the sessions on 6 and 7 September.

  The Counterfactual-Swap Plausibility Audit, record 10.5281/zenodo.22304109 version 1.0.0: the method note, the README, result-run1.json, result-run1R-full-transcripts.json and the record description,STATUS.md, the 4 September 2026 note,research/naming-classification-result.json and research/naming-classification-rerun-2026-09-04.json

• sourceThe readers were said to have judged the published description of the project's research deposit (10.5281/zenodo.22244231).

  Now: They judged the project's own seven-bullet summary of that deposit as held in research/published-record.json on 1 September 2026 (repository commit 909db31, 24 August 2026), not the description published on Zenodo. The summary's first bullet still said six OECD DAC policy markers, a count corrected to eleven the same day it was written (corr-066); no reader in any version challenged it. The material field of result-run1.json always said this correctly.

  Our own error: the method note and the description were written from the design rather than from the result file's own material field, and the two were never diffed before deposit.

  The Counterfactual-Swap Plausibility Audit, record 10.5281/zenodo.22304109 version 1.0.0: the method note, section on the two runs, and the record description

• figureThe material described as a text of 207 words whose three versions differ in exactly seven words.

  Now: 108 words: the title and the seven bullets (115 whitespace-separated tokens when the seven bullet markers are counted; 191 words for the full prompt with the question). The men-and-boys and women-and-girls versions differ at seven positions and share 101 words. The older-persons version replaces ten words with eight, 106 words with 98 shared. No version contains the words male, female, gender or sex. Run 1's readers saw the same title and bullets word for word, plus two label lines ('Title:' and 'What is deposited:') that the run 1R reconstruction omitted; that difference is now recorded in the run 1R result file.

  Our own error: the count was never recomputed from the held arm texts before deposit. Recounted from arms_full_text in result-run1R-full-transcripts.json on 6 September 2026 and again on 7 September; the review of this entry on 7 September also recovered the prompts run 1's readers actually received from the six sessions and compared them word by word with the held text.

  The Counterfactual-Swap Plausibility Audit, record 10.5281/zenodo.22304109 version 1.0.0: the method note and the README (seven of 207 words)

2026-09-03 · 1 correction

• scopeThe audit appendix shipped with the paper states: "CORRECTED: a detail was wrong and the fix is applied in this text." A row-by-row diff of the text against the appendix on 3 September 2026 found the fix actually applied for 13 of the 34 CORRECTED rows. Twenty-one recorded fixes never reached the text, among them the scope of the "Twenty-five funders" header, the 33-offices lodged claim, the section 5 settlement wording, and a reference naming "Ruiz-Grossman and colleagues" for an article that is by Nennstiel and Hudde (European Sociological Review 41(6), doi 10.1093/esr/jcaf021).

  Now: All 21 recorded fixes are applied in version 1.2.0, each re-derived from the primary the audit names and adversarially re-verified before landing. The appendix table itself is unchanged in the deposit, because it is the record of the audit as run; what was false was the introduction's claim that the fixes were applied.

  Our own process error, twice deposited: the audit produced verdicts and recorded fixes, and nothing ever diffed the text against them before either deposit. The diff now runs before any deposit. The one earlier instance caught separately, the MPTF figure, was corrected as corr-097; this entry covers the remaining twenty-one and the false applied-in-this-text claim itself.

  the deposited paper, versions 1.0.0 (10.5281/zenodo.22244232) and 1.1.0 (10.5281/zenodo.22270336), Appendix A introduction and 21 body passages,drafts/paper/the-missing-column.md and paper.html

2026-09-02 · 8 corrections

• sourceThe 79-cents tile attributed the per-beneficiary-dollar circumcision figure to Norway's 2023 rows, and printed only the low end of the range.

  Now: The figure is the global IATI register sweep: of every beneficiary dollar, 79 to 80 cents is circumcision programming, own recount 80.1 (reg-093). Norway's 2023 rows contain zero male-beneficiary projects (reg-091) and cannot produce the figure. The tile now says 79 to 80 cents and names the sweep.

  Transcription slip: the same misattribution was found and fixed in the research paper before deposit, and the article tile kept the older wording. Caught by the post-publication verification pass, rechecked against reg-091 and reg-093 before this entry was written.

  the-counting-machine.html,tellemaskinen.html,data/article27.json,data/article27.nb.json

• figure"94 dated corrections, one against its own front page."

  Now: The sentence now carries no counts, per the repo rule preferring wording that cannot drift: the log is public and dated, and has repeatedly cut against our own front page. At the date of this entry the log held 96 entries, and 38 of them name the front page in their where field.

  The published 'one' was an undercount. The first drafted replacement, 'seven', was computed by a grep for the literal string index.html and was itself an undercount: 31 further entries say 'front page' in words. Caught by the branch re-audit before merge; the wording now points at the log instead of counting it.

  the-counting-machine.html,tellemaskinen.html,data/article27.json,data/article27.nb.json

• figurereg-102: "Largest single funder in 2023 and 2025: the UN's own MPTF Office (USD 61.9M and 70.8M, all non-core; the 2023 amount coincides with the unrelated aid-class figure in reg-092)". The paper: "USD 61.9 million and 70.8 million, all non-core, an amount coincidence with the unrelated aid-class figure in section 3.2".

  Now: USD 51.9 million in 2023, the portal's own 51,942,943 as recorded on 20 August 2026, and 70.8 million in 2025, unchanged. The MPTF Office stays the largest single funder in both years, ahead of Sweden's 51.2 million in 2023. There was no coincidence to note.

  The 61.9 is the world male-directed grant aid class from reg-092, transposed onto the MPTF row; the held portal read, FACTS.md section 31 and corr-075 all carried 51.9 already. The paper's own verification pass on 2 September 2026 recorded this row as CORRECTED with the fix to apply, and the fix was not applied before the register row was written and the paper deposited. Found again the same day on the QA board and rechecked against the held portal file before this entry. The deposited PDF carries 61.9 until a new version with an erratum is issued.

  the public register, reg-102 (js/i18n.js, facts.json),research/mptf-routing.json,drafts/paper/the-missing-column.md and paper.html, the deposited paper (10.5281/zenodo.22244232)

• figure"Climate has five. Disability has one." and the marker tile "climate has five, disability has one".

  Now: The environment has five, two of them for climate. The five Rio markers are biodiversity, climate change mitigation, climate change adaptation, desertification and environment; climate is two of the five, plus an aggregate code that is not a marker. Disability has one.

  A compression of corr-066's own sentence, which named the five environmental markers correctly, that then attached the five to climate. Rechecked against the held codelist, research/oecd-dac-policy-markers.json, and research/dac-marker-count-correction.json before this entry.

  the-headlines.html,overskriftene.html,the-counting-machine.html,tellemaskinen.html,data/article24.json,data/article24.nb.json,data/article27.json,data/article27.nb.json

• scope"94 per cent of children detained in the administration of justice are boys", and on two surfaces "94 per cent of the children detained" with no setting named at all.

  Now: 94 per cent of children detained in justice, conflict and security settings are boys. A/74/136 paragraph 35 gives one figure for the administration of justice and the contexts of armed conflict and national security together, with 67 per cent in migration detention and 56 per cent in institutions; it publishes no justice-only figure.

  The register row reg-082 and the is-it-true page carry the full scope, and the site's own working rule says never to extend the figure beyond the three settings. Attributing the combined figure to one of the three is the same error in the other direction. Every surface was found by searching for the narrowed wording; the paper body had already been fixed by its verification pass on 2 September 2026.

  index.html, the objections section, both languages (js/i18n.js, and the JSON-LD built from it),the-headlines.html,overskriftene.html,the-counting-machine.html,tellemaskinen.html,data/article24.json,data/article24.nb.json,data/article27.json,data/article27.nb.json,drafts/paper/figures/fig1-enact.html

• scope"The whole world's male-directed aid class", "the entire planet's male-beneficiary class", "the world's entire beneficiary class for men in aid": USD 61.9 million in 2023, with no flow qualifier.

  Now: The same figure with its scope: the male-directed class in grant aid. ODA grants are 55.6 per cent of the CRS 2023 universe. The non-grant remainder was scanned the same way and adds about USD 2.1 million, all of it the same circumcision procedure; the class is not hiding there, but the label has to say which flows were counted.

  reg-092 and research/world-empty-class-2023.json both define the class over ODA grants and record the non-grant remainder separately. Dropping the qualifier made the class read as the whole of aid, which our own broader scan puts about 2.1 million higher. The paper's verification pass restored the qualifier in the paper on 2 September 2026; the pages had kept the shorter label.

  follow-the-money.html,folg-pengene.html,the-empty-class.html,den-tomme-klassen.html,the-purpose-clause.html,formaalsparagrafen.html,the-headlines.html,overskriftene.html,the-counting-machine.html,tellemaskinen.html,data/article7, article21, article22, article24 and article27, both editions

• wordingOverture: "the UN's own record has recognised the gap since 2004."

  Now: "the UN's own record has named the barriers since 2004."

  reg-049 supports that the 2004 agreed conclusions acknowledged, in a subordinate clause, that men and boys sometimes face discriminatory barriers, and that nothing was ever assigned to it. It does not support that the UN recognised the missing machinery; that is this site's finding, not the UN's. The line now claims what the source says.

  the-counting-machine.html,tellemaskinen.html,data/article27.json,data/article27.nb.json

• sourcereg-006 published the correct fall, minus 10 per cent from 2023 to 2024, with the source line naming only the OECD blog of 12 March 2026, which itself states 13 per cent on preliminary data.

  Now: The source line names our own published SDMX query on current-price commitments, which recomputes to 10.3 per cent, and keeps the blog for the falling characterisation and its differing preliminary figure. The figure itself was always right and corr-054 documents the deliberate move from the blog's 13 to the query's 10.

  A correct figure attributed to a source that states a different number is the sourcing-error class corr-072 already names. Found by re-checking an external reviewer's flags against our own held recompute.

  facts.json reg-006,sources.html,submission.html,llms-full.txt

2026-09-01 · 6 corrections

• sourceThe card named offices as having been written to that had not been written to. For Norway by age band it said "we have already asked Eurostat FHI WHO". For Denmark, Finland and Iceland it said "we have already asked Eurostat WHO". Four of the six rows carrying an open ask named at least one body we had never contacted, and three of those four named only bodies we had never contacted while omitting the office we actually wrote to.

  Now: The card says "we have already asked and are waiting for a reply" and names nobody. An open ask is published as a country and a breakdown and nothing else, in dataset/timeline/open-asks.csv, and the build refuses to publish a row where fewer than two offices could plausibly have received it, so a recipient cannot be recovered by elimination either.

  The fault was a join. The tier table recorded an open ask against a country, a quantity and a breakdown, and dropped the provider, although every ask in data/open-asks.json records one. The card then filled the sentence from the publishers column, which lists the offices that PUBLISH that breakdown, not the office that received a letter. Those two lists are different and the page presented one as the other. The letters went to national offices: FHI for Norway, DST for Denmark, Statistics Finland, Statistics Iceland, and CSO for Ireland. Eurostat and WHO were named on a public page as recipients of correspondence they never received. Two Ireland rows read correctly, by coincidence rather than by design, because CSO is the only office in both lists there. Keying the ask to its provider fixed the false attribution and immediately created a second fault, because publishing a correct recipient beside a provider name discloses who was written to, and this project publishes only that an ask is open. The rule it breaks is its own, recorded in data/open-asks.json: a specific send is disclosed when a body has not answered and is legally obliged to, not as a matter of course. The published table is therefore aggregated until no single office can be inferred from it, and a guard fails the build if that ever stops being true. Forty four of the countries in this register are covered by exactly one office, so without that guard the next letter would have republished the problem.

  dataset/timeline/tiers.csv,machinery.html, the card that opens when a suicide line is clicked

• finding"the split between helping men and using them is 50.9 to 49.1", and the same sentence in Norwegian. Read plainly, that says the larger share of the money helps men.

  Now: "the split between using them and helping men is 50.9 to 49.1". The numbers are unchanged. The two labels were the wrong way round.

  The register's own classification defines BENEFICIARY as work where the man is the person it is for and the outcome that improves is his, and INSTRUMENT as work where men are engaged to change an outcome for someone else and the outcome belongs to a woman or a child. research/findings.json records the result as 50.9 per cent instrument against 49.1 per cent beneficiary. The article attached 50.9 to helping men, which is the beneficiary side, so it reported the finding backwards in both languages and in the summary line as well as the body. The direction matters more than the decimal: the corrected sentence says a marginal majority of this money uses men rather than helps them, which is the opposite of what was published. classification.json says the same thing independently, recording the published pair as 49.1 and 50.9 in that order and its own reconstruction as 48.3 and 51.7, which puts even more on the instrument side. The error ran in our own favour, which is the specific thing the same file says we will not do: it states that the published figures were left as they were rather than move a number in our own favour by a decimal. Two sentences that give the pair without naming either side were correct and are untouched. Nothing about the finding changes: it is still a dead heat once circumcision programmes are removed, and the caveat that whether that programme belongs to the beneficiary side is genuinely arguable still stands and is still flagged rather than hidden.

  data/article7.json,data/article7.nb.json,follow-the-money article, English and Norwegian, the summary line and the body

• figure"Earmarked state schemes for men's and boys' outcomes found so far: 0", and Norway counted among five countries where every men's row reads none found.

  Now: The row reads 9.3 million kroner: grant scheme DT-0853, social meeting places for boys and young men, administered by IMDi over kap. 671 post 71, frame 9,300,000 in 2026 and 9,000,000 in 2025, its first year. The scheme's stated purpose travels with the row, in IMDi's own words: crime prevention among at-risk boys and young men. Norway moves to the funded group, which becomes six countries against four.

  The scheme sat in the state's own grant register the whole time, established in 2025, a year before the page was published on 21 August 2026. The register's own re-audit of the Norway ledger, run 23 August 2026, surfaced it; on 1 September 2026 it was verified against three publishers before this entry was written: the grant register at tilskudd.no, IMDi's scheme page with the full recipient lists for 2025 and 2026, and Prop. 1 S (2025-2026) AID. On the site's published rank scale a find that disproves a published claim and triggers a dated correction is a meteorite. This one is the register's own, against its own page: the zero row invited disproof with one source, and the invitation was answered from inside.

  data/article23.json,data/article23.nb.json,data/article7.json,data/article7.nb.json,the money in Norway one-pager and Follow the money, English and Norwegian: the columns tables, the ten-country split and the FAQ

• figure"NOK 100.3M to UN Women in 2026", and combined out-of-country support of nearly 700 million, sourced to the government's press release of 13 May 2026.

  Now: The adopted 2026 budget prints 140,300,000 on kap. 164 post 71 (blåboka, confirmed by the forslag column of Prop. 1 S UD). Combined adopted core support to UNFPA and UN Women is 729.9 million, about 730 million.

  The page sourced a press release when the adopted budget existed and disagreed with it by 40 million. Both documents are the government's own: the May press release says the support is held at 2024/2025 levels, the budget adopted in December 2025 raises the UN Women line by 40 million. The budget is the operative document and the register should have read it first. Both figures were re-verified against blåboka and Prop. 1 S UD on 1 September 2026. The discrepancy between the government's own press release and its own adopted budget is left visible on the page, as a fact about the announcement.

  data/article23.json,data/article23.nb.json,data/article7.json,data/article7.nb.json,data/answers-mens-policy-by-country.json,data/stats/cells/funding.jsonl,the out-of-country figures on the Norway money pages and the ten-country answers

• figure"Oslo's scheme 5M" and the row "Oslo municipality, equality and women's liberation grants, annual frame: NOK 5M".

  Now: The allocated frame for 2026 is 4,538,000, the SUM line of byrådens sak 15/26 of 26 February 2026. The three-year series is 4,741,000 (2024), 4,453,000 (2025), 4,538,000 (2026).

  The 5 million was the scheme page's own round number, and the fact row recorded it with a tilde as an approximation. The page then printed it flat, as the frame. The award decisions print exact sums, they were available, and the 2026 decision was verified by refetching the decision document and reading its SUM line on 1 September 2026. An overcount of 462,000 against the operative frame.

  data/article23.json,data/article23.nb.json,FACTS.md,the columns table on the money in Norway one-pager

• figure"Bufdir, family and equality measures, NOK 17.6M frame 2026".

  Now: The operative 2026 frame is 21,600,000: Tildelingsbrev til Bufdir 2026 Tillegg nr. 1, dated 10 March 2026, raises the scheme by 4 million kroner as an extra appropriation for the year.

  Every source the row cited printed 17.6 million at the read date, and two of them, bufdir.no's scheme page and the grant register at tilskudd.no, still print 17.6 million on 1 September 2026. The tildelingsbrev amendment overrides them all and had been in force for five months when the page was published. The lag sits in the state's own registers; the register's error is that it read the registers and not the instruction. The amendment was verified by fetching the Tillegg itself on 1 September 2026.

  data/article23.json,data/article23.nb.json,data/article7.json,data/article7.nb.json,data/stats/cells/funding.jsonl,the neutral-scheme row on both Norway money pages

2026-08-30 · 4 corrections

• sourceSelvmordsrate, menn mot kvinner: 12.4 vs 5.9 per 100k, linked to the WHO age-standardized suicide rates indicator

  Now: Selvmordsrate, menn mot kvinner: 12.4 vs 5.9 per 100k (rå rater), linked to the WHO crude suicide rates indicator

  12.4 and 5.9 are crude rates. The English register said so and cited WHO's crude indicator; the Norwegian register dropped the qualifier and cited the age-standardised indicator, which publishes different values for the same year. A Norwegian reader following the link to check the figure would not have found it, and would have been right to conclude the number was unsupported. Crude and age-standardised rates are different measurements and this site's own published rule is that they are not interchangeable, so citing one for the other is the exact error the register exists to catch. Found by tools/check-language-parity.mjs, which had been reporting the pair as a discrepancy without anyone reading which of the two editions was wrong.

  js/i18n.js (nb register, reg-014),index.html (nb)

• sourceOne register row carried two different findings from two different sources: the male share of US deaths of despair, compiled by AIBM, and the concept and mortality series, which are Case and Deaton's. The English edition linked to Case and Deaton at Brookings while stating a share that is not their figure. The Norwegian edition linked to AIBM.

  Now: Split. reg-067 carries the AIBM-compiled male share and links to AIBM in both editions. reg-081 carries Case and Deaton's own finding and links to Brookings in both editions.

  A reader clicking the English source to check nearly three in four was sent to a paper that does not state it. Worse, Case and Deaton report that deaths of despair are rising in parallel for men and women without a high school degree, so citing them beneath a mostly male headline risked implying they found a male-specific trend, which they did not. The two findings are compatible, a share and a direction, but they are not the same claim and must not rest on one row or one link. The parallel-rise finding is now registered in its own right rather than left out because it complicates the headline.

  js/i18n.js (reg-067, both editions),facts.json,index.html,index.html (nb)

• figurePaid 2024 and 2025 combined, SEK 48M, for the state grant for women's and girls' organising under regulation 2023 168, with a footnote giving 20 3M in 2024 plus 18 9M in 2025

  Now: Three separate lines: the appropriation, SEK 48M per year; organisationsbidrag paid SEK 25.97M and 27.37M for grant years 2024 and 2025 to 67 associations; verksamhetsbidrag paid SEK 20.33M and 18.91M in the same decision round to 17 organisations

  The row did not close on its own arithmetic before any source was opened: the two figures in its own footnote sum to 39.2M, not 48M. The primary says why. Jamstalldhetsmyndigheten Rapport 2026:14 states "Anslaget har hojts till 48 miljoner kronor per ar", so 48M is the yearly appropriation and never was a two year total, and the 20.3M and 18.9M in the footnote are verksamhetsbidrag alone, one of two grant streams. The row therefore mislabelled the period and, by quoting one stream, understated what was paid. The correction moves the figure in the direction that strengthens the article's own argument, which is worth stating plainly: we were under-claiming. An appropriation and a payment are different measures and are now on different lines rather than added, which is the same rule that keeps counts apart from rates elsewhere in this project. A figure of 5 999 tkr that appeared in the working notes could NOT be found in the report and was not used.

  data/article7.json,follow-the-money article (both languages)

• method72 projections published on 30 August 2026, each an ordinary least squares line fitted to the last ten observations and run out to a fixed year 2030

  Now: Those 72 are retired and kept in full under a "retired" list in the same file, with the date and the reason. Every live projection is now a damped trend, capped at five years past the last observation, and carries a band equal to the method's own measured error at that horizon

  The method was backtested after publication and did not beat doing nothing. Rolling origin over every series in the register, 8940 forecasts from 2074 origins, scored as mean absolute error divided by each series' own mean year-on-year move so that series in years, per cent and counts can be pooled: the published method scored 1.838 and carrying the last observed value forward unchanged scored 1.737. The published method lost at every one of the five horizons. 32 of the 72 had an r squared below 0.5, meaning the line did not fit even its own past. A damped trend scored 1.482 and won at every horizon. Nothing had been scored against any of the 72: they were published and superseded within the same day, so no reader had yet relied on one. The projections were never fabricated and each names the years it was fitted on, so any of them can still be reproduced by hand from the retired record. What was wrong was narrower: we published a line that implied more knowledge than the method had.

  dataset/timeline/forecast.json,machinery.html, the projected years on every outcome chart

2026-08-29 · 1 correction

• wording232 rows, resolving to 71 distinct projects

  Now: 232 rows, resolving to 71 distinct activity texts

  71 is the count after deduplicating by activity text. The 232 rows carry 232 distinct microdata ids, so nothing is folded at project level, and the sweep file names its own field male_naming_distinct_texts. The stronger word was introduced downstream of the data that does not support it. This entry was made on 28 August and withdrawn the same day on the grounds that the text was merely stale and the fix had already landed. Neither was true: the value never moved, and the wording was still live on the stat tile in both languages and in classification.json for another day. It is restored because withdrawing it also erased the only record that a DOI deposit carries the unsupported claim, and a deposit cannot be edited. Zenodo 1.3.0 keeps the old wording in data/classification.json and data/timeline/markers.json permanently; this entry is the erratum for it.

  the-empty-class.html,den-tomme-klassen.html,classification.json,Zenodo deposit 1.3.0

2026-08-28 · 6 corrections

• figureOne 70-item CRS++ field list serves all eleven markers, and none is a beneficiary, target-group, destination or results field.

  Now: One 70-item CRS++ field list serves all eleven markers, and none is a beneficiary, target-group, sex or age field.

  The destination half of that sentence was false and this project's own data refutes it. CRS field 17 is Geographical target area, and the eleven marker-splits files published here are recipient splits pulled from that dataflow, so the register asserted a field does not exist while eleven of its own files were built from it. The verified finding, against Annex 4 of the Converged Statistical Reporting Directives DCD/DAC(2026)14/ADD1/FINAL approved by the DAC on 28 May 2026, concerns who a project is for, not where the money lands. The error shipped inside the published 1.3.0 deposit.

  FACTS.md,research/rio-marker-pattern.json,research/zenodo-deposit/data/rio-marker-pattern.json,Zenodo deposit 1.3.0

• figureThe gap has roughly tripled.

  Now: The gap has more than doubled, by a factor of about 2.6.

  EU tertiary attainment aged 25 to 34: the gap is 11.3 percentage points in 2024 against 4.3 in 2002. That is 2.63 times, which is more than double and is not triple. The underlying Eurostat figures verified exactly against edat_lfse_03; only our characterisation of them was inflated, and it was inflated in our own favour. The recount of 26 August recorded the error and the register was told to refuse the word, but the two evidence files still carried it and one of them shipped inside the published 1.3.0 deposit.

  research/evidence-of-need.json,research/zenodo-deposit/data/evidence-of-need.json,Zenodo deposit 1.3.0

• sourceThe research corpus is archived with a DOI: 10.5281/zenodo.22131703 (version 1.3.0 published 27 August 2026).

  Now: The research corpus is archived with a DOI: 10.5281/zenodo.22073333 (all versions; version 1.3.0 is 10.5281/zenodo.22141707, published 28 August 2026).

  Version 1.3.0 was first uploaded on 27 August as a new upload rather than as a new version, so it received its own concept DOI, 22131703, outside the deposit's real version chain. The correct 1.3.0 was published on 28 August inside the 22073333 family and Zenodo has been asked to withdraw the orphan. Both pages were still citing the orphan, which would have become a dead citation on withdrawal.

  about.html,README.md

• figureWe checked five countries. Not one publishes the answer.

  Now: We checked four countries. Not one publishes the answer.

  The register's own row, reg-054, lists four jurisdictions: Norway, Denmark, Sweden, and England and Wales, which is one jurisdiction. The beat counted five. Both language editions carried it.

  js/i18n.en.js,js/i18n.nb.js

• scopeThree new markers were added in 2018 alone.

  Now: Three markers carried no money at all before 2018.

  research/marker-series.json states that first year with data is an upper bound on adoption and must never be described as an adoption date. The adoption dates would need the DAC reporting directives, which have not been read for this and are not claimed.

  gap.html

• scope7,270 aid rows read one by one

  Now: 7,270 aid rows swept, 71 read one by one

  The 7,270 rows were swept by a programmatic matcher. Only the 232 rows it returned, resolving to 71 distinct activity texts, were read and classified by hand. The beat transferred the hand-reading onto the whole corpus.

  gap.html

2026-08-27 · 3 corrections

• figureBoth money pages called Norway the fifth-largest state funder of UN Women in 2023, 'femte største statlige giver' in the Norwegian edition.

  Now: Fourth-largest state funder. The register's own capture of the 2023 table has the UN Multi-Partner Trust Fund Office first at 51.9 million dollars, then Sweden 51.2, Germany 39.8, Finland 32.0, Norway 30.9. Three states sit above Norway; the MPTF Office is a UN pooled fund, not a member state, which the register row itself flags. Fifth in the table overall, fourth among states.

  The page wording combined the table rank with the word 'state' and the register never recorded a rank at all, only the 30.9 million contribution. Found on 27 August 2026 in an internal adversarial verification pass run before the figures were reused, checking every page claim against its register row. The contribution figure itself was and is correct.

  pengene-i-norge.html and the-money-in-norway.html, via data/article23.json and data/article23.nb.json

• figureBoth money pages and the aid article said Reform is one of the four likestillingssentre paid over kap. 351 post 73.

  Now: Five, not four. The fifth centre, for region vest, was established in 2024. Prop. 1 S states that the five centres together received 27.5 million kroner in base grants from the ministry in 2024, and the 2025 allocation row shows five centres applying. Reform is one of the five. The post 73 amount of 29,450,000 kroner for 2026 stands unchanged.

  The count came from the earlier budget reading and was not refreshed when the site's own later research, held since 23 August 2026, recorded the fifth centre. Found on 27 August 2026 in the same internal verification pass, when the newest held source contradicted the published count.

  pengene-i-norge.html, the-money-in-norway.html, folg-pengene.html and follow-the-money.html, via data/article23 and data/article7 in both languages

• sourceThe purpose clause article said the committee opened written input on 21 August, as if the opening date came from Stortinget.

  Now: The date is this site's own reading date. The sentence now says written input stood open with the 28 September deadline when the case was read on 21 August. The Stortinget case page confirms the deadline and that input is open to everyone; it does not publish an opening date.

  A sighting was phrased as a source fact. Found on 27 August 2026 in the same verification pass, by checking the live case page against the article sentence.

  formaalsparagrafen.html and the-purpose-clause.html, via data/article22.json and data/article22.nb.json

2026-08-26 · 6 corrections

• figureUSD 70.8 billion a year of aid marked for gender equality, 2023-24 average, on the front page counter and across seventeen other strings in both languages.

  Now: USD 71.8 billion a year, the figure OECD's own SDMX endpoint returns.

  Not wrong when written, and known since 19 August. The 70.8 came from Focus 2030, a compiler working from an earlier vintage. Checking it directly at OECD on 19 August 2026 gave 71.8bn for the same key shape on dataflow DAC_EC, recipient DPGC, all sectors, commitments at current prices, with the marker score varied across its own position. Two of the three cross-checks matched exactly at the same time, principal at USD 6.0bn and share of screened bilateral aid at 48.2 per cent, which is what showed the series was right and the gap was vintage rather than scope. The register recorded the discrepancy that day and deliberately left it published, because moving it touches eighteen strings across two languages and should be one considered edit rather than a side effect of another. This is that edit. Two places keep 70.8 and are not errors: the Multi-Partner Trust Fund Office figure of USD 70.8 million in the funding article, which is a different quantity, and the earlier entries in this log, which record what was said at the time and must not be rewritten.

  index.html, is-it-true.html, sources.html, submission.html, watch.html, follow-the-money.html, the-one-way-glass.html, vi-spurte.html, js/i18n.js, data/answers-is-it-true.json, data/article7.json, data/article15.json and their Norwegian editions

• scopeAbout section 06 stated the costs of running this were a domain, hosting at a few hundred kroner a year, and unpaid time.

  Now: It runs on paid monthly tooling: the domain and its mail routing, Cloudflare for the pages and the API, GitHub for the repository and the 57 step build, and an AI subscription used for harvesting and cross-checking. All paid personally, on top of the unpaid research time.

  True when written and outdated since, which on a section about money is the same problem as being wrong. Caught while building the contributions register, where the same understatement was about to be published a second time. The tooling is now named on both pages, together with the rule that no figure rests on it: harvesting is not verification, and every figure was checked against a primary source by a person.

  about.html, contribute.html

• figureThree story beats, in English and Norwegian, still carried the pre-correction prison figures after the 19 August entry declared every instance moved: the story said seventy rules were written for the six per cent who are women and that nobody wrote the supplement for the ninety-four, and the Norwegian number tile still read 94%.

  Now: The beats read seven per cent and ninety-three in both languages, matching the corrected lead and UNODC Prison Matters 2026, reference year 2024. The FACTS.md canonical row, which still said 94% citing Prison Matters 2025 and would have reseeded the old figure into future copy, is updated with a supersession marker.

  The 18-19 August sweep fixed the number tile in the English story and missed the two prose beats beside it, and missed all three in the Norwegian table. Found by the second pass of a full audit. A correction entry that claims every instance moved is itself a claim, and this one was wrong; that is worth logging as plainly as the figure.

  js/i18n.js (English and Norwegian story beats), js/i18n.nb.js and llms-full.txt via rebuild, FACTS.md

• sourceAfter the 26 August correction moved the gender-marked aid figure from 70.8 to 71.8 billion, llms.txt still told machines 70.8, and the register's source label on the corrected figure still credited Focus 2030, the compiler whose earlier vintage supplied the wrong number.

  Now: llms.txt says 71.8, and the source label reads OECD SDMX endpoint with a link to the exact keyless query that returns the figure, which is where the correction said the number now comes from.

  The sweep's where-list covered the pages and missed the hand-typed llms.txt string, and nobody updated the label when the source of truth changed underneath it. A correct figure attributed to the source of the wrong one is a sourcing error even when the digits are right. Found by the full audit; the scope was narrowed by hand after a first proposal to sweep 70.8 to 71.8 numerically, which would have corrupted three legitimate 70.8s elsewhere.

  llms.txt via tools/build-llms.mjs, js/i18n.js registers in both languages, facts.json, sources.html and llms-full.txt via rebuild

• figureTwo answer pages said 10 OECD policy markers exist, citing a dataflow that holds five of them, after the 24 August correction established eleven. The ten was eleven minus the gender marker itself, a denominator nobody had written down, and the eleven-marker page cited the same five-marker dataflow for its own count.

  Now: Both pages say 11, and all three marker claims cite the number to the binding text that states it in words: the Converged Statistical Reporting Directives, DCD/DAC(2026)14/FINAL chapter 4, which is what our own marker-count correction file instructs. The visualisation's Gender label is a display name for the gender equality marker, one of the eleven, and is unchanged.

  Found by the full audit as a contradiction between pages; held for a human decision because ten-of-eleven-excluding-gender was a possible deliberate denominator; decided to state eleven. An undocumented denominator that has to be reverse-engineered is a wrong figure in practice even when the arithmetic behind it was defensible.

  gendered-and-raced.html, objections-to-a-un-for-men.html, is-it-true.html, via data/answers-gendered-and-raced.json, data/answers-objections-to-a-un-for-m.json, data/answers-is-it-true.json

• scopeThe headlines article opened, in English and Norwegian, with 'We read the whole UN Women publication library. 625 documents', after the complete census of 21 August found the library lists 971, meaning the earlier corpus covered 625 and missed 343 from 1997 to 2019. FACTS.md recorded this correction when the full census ran; the public log never did.

  Now: The article opens 'We read 625 of the 971 documents in the UN Women publication library' in both languages. The figures inside the article describe the 625-document corpus and remain correct for that corpus; the complete census, with the corrected 7.49 to 1 ratio that replaced the roughly 10.5 to 1 the partial corpus implied, is published separately.

  A correction documented internally but absent from the public log is half a correction. Found by the full audit's translations pass, which also found the Norwegian edition carried the same overstatement.

  checked-in-public.html and etterprovd-i-full-offentlighet.html, via data/article25.json and data/article25.nb.json

2026-08-24 · 3 corrections

• figureThe OECD DAC described as operating six policy markers, with "a seventh, for men and boys" said not to exist.

  Now: Eleven policy markers, with a twelfth for men and boys not existing. The chart still shows six, and now says so.

  The DAC publishes its markers across three OECD SDMX dataflows. Our harvest queried two of them and never touched DSD_RIOMRKR, which carries five Rio markers for biodiversity, climate mitigation, climate adaptation and desertification. Six was our own undercount, not the DAC's. Eleven is the harder number and correcting it strengthens the page. The 65,069 million dollar gender figure is unaffected, having been queried from the gender dataflow directly.

  follow-the-money.html, folg-pengene.html

• figureWHO Global Health Observatory listed as 3,090 indicators, women and girls 57, men and boys 6, with no disaggregation of the 57.

  Now: The same counts, now followed by the split: 43 of the 57 have no biological male counterpart, and removing those plus the one matched pair leaves 13 against 5.

  The raw pair invites a reading of nine and a half to one. Recounting the catalogue on 24 August 2026 showed that 43 of the 57 concern pregnancy, childbirth, maternal mortality, cervical and uterine conditions, contraception and midwifery, which could not have a male version. The honest comparison is about two and a half to one. The finding survives and is sharper for it, because all five remaining male indicators are infectious disease surveillance rather than measures of men's own health.

  the-empty-room.html and its Norwegian edition

• figureThe suicide interval given as 63.9086 seconds.

  Now: 63.9084 seconds.

  31,536,000 divided by 493,456 is 63.9084, not 63.9086. Nothing visible changes, since it rounds to 64 in prose and to 63.9 at one decimal either way. It is corrected because this page argues at length that "one every 64.0 seconds" is arithmetically wrong, and a page that strict about a first decimal should be right in its fourth. The earlier corrections-log entries that quote 63.9086 are left as written, because a corrections log records what was said at the time.

  suicide.html

2026-08-22 · 5 corrections

• figureEngland's men's health strategy money described as "roughly one two-hundredth of what one Nordic country sends to two UN gender partners in a single year".

  Now: "Roughly one forty-sixth, annualised and converted at Norges Bank's rate."

  The ratio divided kroner by pounds without converting either. At Norges Bank's official rate for 21 August 2026, 12.6853 kroner to the pound, NOK 700 million a year is GBP 55.2 million; GBP 3.6 million spread over three years is 1.2 million a year, so the honest ratio is about one to forty-six. Found in an expert language review commissioned by this site. A first correction published the same day said one to forty-four, computed on an assumed rate rather than a quoted one; that is why this entry names the rate, its source and its date.

  follow-the-money.html, folg-pengene.html

• quotationThe 2004 agreed conclusions paraphrased as stating that "men and boys face discriminatory barriers and practices".

  Now: "Men and boys sometimes face discriminatory barriers and practices", which is the endorsed text's own word.

  The endorsed paragraph, ECOSOC agreed conclusions 2004/11, reads "while men and boys sometimes face discriminatory barriers and practices". Dropping one qualifier made the source say something slightly stronger than it says. The register held the verbatim sentence all along; the page had drifted from it.

  un-for-men-2004.html, un-women-on-men.html and their Norwegian editions

• figureThe search capture linked as "four queries, thirty-two results".

  Now: "Six queries, forty-eight results", which is what the published capture file contains.

  The label undercounted the record it points at. The capture itself was complete; the description of it was not.

  what-the-un-does-for-men.html and its Norwegian edition

• wordingThe method described as a "pre-registered hypothesis".

  Now: "The hypothesis written down and committed to the public repository before reading."

  Pre-registration means lodging a protocol in an independent timestamped registry. Committing rules to our own repository is not that, however honestly it was done, and a methodologist checking for a registry entry would have found none. What happened is still checkable, and is now described exactly.

  the-empty-class.html, den-tomme-klassen.html

• wordingThe DAC marker called "maternal and newborn health" and described as being "about one sex".

  Now: The marker's official name, reproductive, maternal, newborn and child health, with the sex-specificity claim resting on its maternal core.

  Newborn and child health covers both sexes, so the shorter label overstated. The argument that a sex-specific objective already has a box survives on the maternal component alone.

  follow-the-money.html and its Norwegian edition

2026-08-20 · 1 correction

• figureThe empty room stated the five-registry census totalled 66,900 indicator and dataset names, in the summary and the closing section.

  Now: 47,100, which is what 3,090 + 713 + 12,227 + 1,537 + 29,533 actually sums to. The per-registry counts were correct throughout; the headline total was mis-added.

  Caught within hours while drafting social posts from the article, because writing a number a second time is a recount. The dedicated-attention counts, 1,220 female-named to 204 male-named across the five registries, were unaffected.

  the-empty-room.html

2026-08-19 · 6 corrections

• figure“The flow is now falling, down 13% from 2023 to 2024”, on the front page and in the register, citing the OECD blog of 12 March 2026. Also “OECD DAC members' commitments” as the label on the 70.8 billion figure.

  Now: “down 10% from 2023 to 2024”, and “OECD DAC members and EU institutions, commitments”.

  Not wrong when written. The OECD blog says, in its own words, that bilateral allocable ODA with gender equality objectives “dropped by 13% from 2023 to 2024”, and the site reported that accurately. Re-queried against OECD's own dataset on 19 August 2026, the same series gives a fall of 10.3% in current prices and 11.6% in constant prices, which is what revision of preliminary 2024 data looks like. We publish the figure our own query returns and keep the blog cited for the characterisation. The aggregate label was also imprecise: the figure covers DAC members and EU institutions together, which is what OECD's DAC_EC aggregate is, not DAC members alone. Two of the three headline numbers were confirmed exactly at the primary source in the same pass: 6.0 billion principal, and 48.2 per cent of screened bilateral aid. One caveat we are recording against ourselves: on a disbursements basis rather than commitments, the same flow rose about 18 per cent from 2023 to 2024. Commitments is the standard basis for gender marker analysis and is what the OECD blog and this site both use, but a reader should know the other basis exists and points the other way.

  index.html, sources.html, submission.html, facts.json, FACTS.md

• figureEvery figure in the marker explorer on the money article, published earlier today: gender-marked aid of $3,480m for 2022, governance $3,389m, disability $264m, and “28 of 32 donors” marking aid for gender equality.

  Now: Gender-marked aid of $65,069m for 2022, governance $40,154m, disability $16,001m, and 33 of 33 donors marking aid for gender equality. Every donor does.

  The OECD query was scoped to the wrong recipient. In OECD's CL_AREA_ORG codelist the code “A” is America, not “all recipients”, and nothing in the key says so. The pull ran, returned plausible figures, and cross-checked cleanly against OECD's own DAC and EU aggregate to the decimal, because a regional slice is internally consistent with itself. What caught it was comparing the total against a figure from outside that dataset: this site already publishes about $70.8bn a year of gender-marked aid, and $3.5bn could not be the same quantity. The correct code is DPGC, developing countries, which gives $65bn for 2022 and sits where it should against $70.8bn for 2023-24. The correction makes the campaign's own point harder rather than easier: it is not 28 of 32 donors who use the gender marker, it is all 33 of them, and the instrument men and boys lack is one every single donor uses.

  follow-the-money.html, folg-pengene.html, FACTS.md, research/donor-markers-2022.json, research/marker-totals-2022.json

• findingOn sixteen-populations.html and seksten-grupper.html, published earlier the same day: “We have not read the European Programme of Work. Until we have, the honest statement is that we found no successor men's strategy and no new reporting duty.”

  Now: We read it. The successor is the Second European Programme of Work 2026 to 2030, adopted unanimously by all 53 member states in October 2025, and it cites the 2018 men's health strategy in its reference list behind a commitment on mid-life interventions addressing the differences in women's and men's health. The article now carries that in the body.

  The original sentence was true when written and was written to be falsifiable on purpose. It was falsified within hours, by us, and it went the way that weakens our own argument. The strategy was not dropped; it is carried forward as evidence in a unanimously adopted programme. What survives is narrower and is now stated as such: EPW2 is a programme of work rather than a men's strategy, the 2023 reporting duty was not replaced, and the asymmetry has moved from the adult sections to the children's, where girls have a named Spotlight section and boys appear three times in 144 pages, every time as the line against which girls are measured.

  sixteen-populations.html, seksten-grupper.html, research/epw2-2026-2030.json, FACTS.md 25

• findingOn the-word-that-is-not-there.html and ordet-som-ikke-star-der.html, published earlier the same day: three UN documents carrying the same signature, presented under the heading "Three chains, not one". The article hedged that three documents is not an institution, but a reader could reasonably have taken it as a UN-wide habit.

  Now: A seventh section carries the opposite evidence. The Human Rights Committee, a different treaty body, names girls and boys together as a matter of routine. Its general comment No. 35 of 2014, on the article that governs detention, lists who it protects and begins the list with girls and boys. The closing and one refused claim were rewritten to match.

  We ran the same count on a fourth chain expecting the same shape, and got the opposite. There is no UN-wide habit of omitting boys, and saying there was would be refuted by our own evidence. The narrower finding is stronger: it is a difference between two committees writing about the same children, and the one that omitted the word was the one writing specifically about the system those children are 94 per cent of. It also changes the ask from please add a name to please use the sentence another UN committee already wrote.

  the-word-that-is-not-there.html, ordet-som-ikke-star-der.html, research/ccpr-general-comments-census.json

• figureThe correction logged on 18 August, moving the male share of the world prison population from 94 to almost 93 per cent, was applied to one sentence and left everywhere else. Still live afterwards: the front page displayed 94% as a headline number; violence.html showed a 94% figure directly above a claim reading 93 per cent; its lead said “ninety-four in every hundred” and “the other six”; the Norwegian source table still read 94 % and UNODC 2025 throughout; and the stored quotation carried the new percentage with the old year, attributed to the old report, so it matched neither edition.

  Now: Every instance moved to UNODC Prison Matters 2026, reference year 2024, verbatim: “The vast majority of prisoners worldwide are men, representing almost 93 per cent of the prison population or nearly 11 million prisoners in 2024.” The lead now reads ninety-three and the other seven, which the same report supports: the share of women stayed below 7 per cent from 2014 to 2024. The source URL, which was missing, is recorded and its SHA-256 matches the copy held locally.

  The mangled quotation is the serious part. Editing a number inside quotation marks while leaving the year and the citation produced a sentence that appears in no UN document. Nobody did it deliberately; a correction was applied to the prose and the surrounding fields were not re-read. This site's whole method is that a reader can open the source and find the sentence, and for a day that would have failed. Found by an automated sweep for the same quantity published with two close values, written after a research file was found contradicting itself.

  index.html, violence.html, sources.html, facts.json, data/violence.json, js/i18n.js (English and Norwegian source tables), data/evidence-manual.json

• quotationOn the-word-that-is-not-there.html and its Norwegian edition: all three blockquotes truncated mid-sentence and closed with a period as if complete, one pair under a standfirst reading “Here they are in full”. GC24 paragraph 40 stopped at “children with disabilities”, dropping the accommodation clause. GC35 paragraph 3 stopped at “migrant workers”, dropping “persons convicted of crime, and persons who have engaged in terrorist activity”, the two most detention-relevant entries in the list, in an article about selective omission. GC10's sentence stopped at “the girl child”, dropping “e.g. in relation to prior abuse and special health needs”. GC10's own discrimination passage, paragraph 6, was promised by the standfirst and not shown. The A/74/136 sentence was cut at its semicolon. The closing line said four mentions of girls without the whole-word qualifier, though the singular girl child appears twice more. And the same truncation had run on un-for-men-2004.html since first publication: the 2004 paragraph-2 quotation stopped at “in many capacities”, mid-sentence.

  Now: Every quotation extended to the end of its sentence, verbatim from the held PDFs. GC10 paragraph 6 is quoted in full alongside GC24 paragraph 40, so the standfirst is now true rather than deleted. The closing tally reads six to zero on the widest count, with the counting method beside it. The 2004 sentence is complete in all four editions that quote it.

  Every number in the article was right and every quotation was wrong in the same direction. The counts were checked mechanically; the quotes were not, because nothing checks them. All five defects came from a reader within hours of publication, all five verified against the held documents, and the finding survives every one. An article whose authority is that we count exactly and quote exactly has to do the second one too. An audit of every quotation on the site against the held corpus found the rest sound: 22 verified clean, and the pattern was confined to the pages named here.

  the-word-that-is-not-there.html, ordet-som-ikke-star-der.html, un-for-men-2004.html, fn-for-menn-2004.html, who-runs-the-un.html, hvem-styrer-fn.html, FACTS.md 27, research/crc-general-comments-census.json, research/ccpr-general-comments-census.json, drafts/legal-brief-A74136.md, drafts/who-to-ask.md

2026-08-18 · 4 corrections

• figure“94 per cent of the world's prison population is male”, quoting UNODC Prison Matters 2025.

  Now: “Almost 93 per cent”, quoting UNODC Prison Matters 2026 for reference year 2024: “The vast majority of prisoners worldwide are men, representing almost 93 per cent of the prison population or nearly 11 million prisoners in 2024.”

  Not an error when written; the source published a new edition and the figure moved. That is the failure mode a citation with a year in it is supposed to prevent and it still got past us, because nothing re-reads a source after the first time. Found by a research review, then verified here by downloading the 2026 edition and reading the sentence. The local evidence archive now holds both editions, so the next move in this figure will be visible rather than invisible.

  violence.html, sources.html, facts.json

• scope“No UN entity for men and boys exists”, stated flatly in the register and on the answers page.

  Now: “No global UN entity dedicated to men and boys”, with the two exceptions set out in full wherever the claim appears: the WHO Regional Committee for Europe adopted a Strategy on the health and well-being of men by resolution EUR/RC68/R4 on 19 September 2018, two years after a women's strategy, with an owning programme and a reporting duty that closed with a final report in 2023; and Security Council resolution 2467 (2019), paragraph 32, recognises men and boys as targets of sexual violence in conflict and urges states to protect male survivors.

  The old form was refuted by two documents this site had not found. An adversarial review was commissioned to attack the campaign's own claims and this is what it returned, so both were then verified here from the primary documents rather than taken on report: the WHO strategy PDF carries “Annex 1 Resolution EUR/RC68/R4” and the operative words “ADOPTS the strategy on the health and well-being of men in the WHO European Region”, and the Security Council text carries paragraph 32 verbatim. Both are held in the local archive. A regional committee of 53 member states adopting a men's health strategy by resolution is a mandate by any ordinary meaning of the word, and calling a constitutional organ of a UN specialised agency “not a UN entity” would not survive scrutiny. The narrowed claim is what the evidence supports and it is also the better argument: the thing this campaign asks for has been adopted once, at regional level, and then allowed to lapse.

  sources.html, facts.json, is-there-a-un-for-men.html

• framing“Father and paternity appear zero times across the 234 SDG indicators”, printed on the 2004 article, the answers page and the fatherhood register without saying what the comparable words for mothers do.

  Now: The same count, followed by the rest of it: mother and maternity are zero too, the framework carries no parental vocabulary at all, and “maternal” appears twice, both times in maternal mortality. Where the difference does show is in who is named as a group: women 37 to men 8, every one of the eight inside a pair, and girls 11 to boys 2.

  The figure was correct and the impression it left was not. Standing alone it invites the reader to conclude that the framework tracks mothers and overlooks fathers, and it does neither. A hostile reader would say the search terms were chosen because they return zero, and on this line they would have been right. Found by an adversarial review commissioned to attack this site's own claims, and confirmed by recounting the official PDF here: father 0, paternity 0, paternal 0, mother 0, mothers 0, maternity 0, maternal 2. The corrected version is a stronger finding, because a framework with no word for a parent is a more interesting fact than a framework that forgot fathers.

  un-for-men-2004.html, fn-for-menn-2004.html, is-there-a-un-for-men.html, fatherhood.html, sources.html

• sourceThe register entry “The UN has known since 2004” quoted the endorsed wording, “while men and boys sometimes face discriminatory barriers and practices”, while citing and linking the advance unedited version at csw48/ac-men-auv.pdf.

  Now: The same quotation, cited and linked to the text it is actually in: the agreed conclusions as endorsed by ECOSOC. The entry also states that two texts exist and gives the advance wording.

  Two texts of paragraph 2 exist and this site found and published that difference itself on 16 August. The register then paired one text's words with the other text's link, so anyone following the citation and searching for the quoted sentence would not have found it. The article's own “Two texts of the same paragraph” block was correct throughout; the error was in the machine-readable register, which is the file a researcher is most likely to check mechanically. Found by an adversarial review of this site's claims.

  sources.html, facts.json

2026-08-17 · 8 corrections

• sourceThe article reproduced a cropped Agderposten news photograph of the MannsForum stand.

  Now: The photograph is removed. The headline is quoted with attribution, and the caption says plainly that the picture is not reproduced and why.

  A newspaper photograph is the newspaper's work. Quoting a headline is ordinary citation; republishing a press photo on a campaign page is not, and a site that asks institutions to respect what they have written down does not get to help itself to someone else's picture.

  un-for-men-2004.html, fn-for-menn-2004.html

• scope“Whoever it was, they were working through the tents nobody would defend”, alongside a claim that specific named stands were hit.

  Now: No motive is offered and no other organisation is named. The theft total is attributed to Agderposten, and the text says outright that the author does not know who did it or why.

  Attributing a motive to an unidentified person is speculation, and this project's own rule is actions, not motives. The rule binds us first. Naming other organisations as victims without a source was an unsourced factual claim about third parties.

  un-for-men-2004.html section 1, fn-for-menn-2004.html

• figure“the most-read thing that subreddit has carried in a fortnight”.

  Now: Only the post's own 15,000 views, with a note that Reddit shows view counts to a post's author and to nobody else.

  The comparison was not checkable by the person making it. Reddit does not expose other posts' view counts, so a ranking against them could not have been observed. A claim that sounds checkable and is not is worse than no claim.

  un-for-men-2004.html section 4, fn-for-menn-2004.html

• scopeParagraph 2 described as “the only place the United Nations has written down that men and boys face discriminatory barriers and practices”.

  Now: Described as where the UN wrote it down, adopted by consensus and endorsed by ECOSOC, with no entity, focal point or indicator ever assigned to it.

  An absolute negative across the whole UN corpus cannot be proven and a single counter-example would destroy it. The register's own framing is recognised and never assigned, which is the claim the evidence actually supports.

  un-for-men-2004.html section 5, fn-for-menn-2004.html

• scopeThe Gemini exchange was introduced as “a reader did the experiment properly, by accident”, and closed with “this one was tested by accident”.

  Now: It says the author ran the test himself and deliberately, that nobody stumbled onto it and nobody independent ran it, and that he built the least friendly reader he could reach and pointed it at his own work.

  There was no reader. The author wrote the hostile prompt and ran it once, which the source notes had said all along, so the piece contradicted itself within one section. Implying an independent third party had produced the result would have made it far stronger evidence than it is, and would have been the single easiest thing to catch us on. The accurate version is also the better one: the test was built to fail and did not.

  un-for-men-2004.html section 4, fn-for-menn-2004.html

• scopeNOU 2024:8 described as “the most thorough public inquiry into men’s equality any country has run”.

  Now: Described as a public commission with a full mandate that reported 319 pages and 35 measures, with the note that a broad equality commission of this kind is rarer than a men’s health policy, that the author has not found another, and that this is not the same as saying there is none.

  A superlative across every country cannot be checked, and the register itself documents Ireland 2008, Brazil 2009, Australia 2010 and England 2025. It was published in the same session in which three other absolute claims were removed from this article for the same reason, which is how the class of error survives: by being introduced again while the last one is being fixed.

  un-for-men-2004.html section 2, fn-for-menn-2004.html

• scope“no source that contradicts the record, because the record is the only place it exists”.

  Now: “in a forum where an answer without a citation is removed, no citation that contradicted the record came back”.

  The old form claimed our own register is the only place this information exists anywhere, which is an absolute negative about every source in the world and cannot be supported. What actually happened is narrower and still worth reporting: in a subreddit that requires citations, no contradicting citation arrived.

  un-for-men-2004.html section 4, fn-for-menn-2004.html

• sourceThe 2004 agreed conclusions were repeatedly described as “adopted by consensus”, on the published article in both languages, in the submission to the UN system and on the homepage.

  Now: “Adopted at the Commission's forty-eighth session. Endorsed by ECOSOC.”, which is what the document says about itself.

  The word consensus does not appear in the document. It was searched and returns zero. Agreed conclusions at this Commission are negotiated texts and are commonly adopted without a vote, so the claim is probably true as a matter of practice, but this site's rule is that a claim carries a primary source and this one carried an assumption. The document's own opening sentence is stronger anyway because a reader can check it in ten seconds: the Economic and Social Council endorses conclusions adopted by the Commission at its forty-eighth session. Found while fact-checking a sentence written for a new article, which turned out to have been inherited from the published one.

  un-for-men-2004.html, fn-for-menn-2004.html, submission.html, index.html

2026-08-16 · 1 correction

• quotationParagraph 2 of the 2004 agreed conclusions was quoted as “men and boys, while some themselves face discriminatory barriers and practices, can and do make contributions to gender equality in their many capacities”.

  Now: “The Commission recognizes that while men and boys sometimes face discriminatory barriers and practices, they can and do make contributions to gender equality in many capacities.”

  Two texts of this paragraph are published on un.org. We quoted the advance unedited version, which is headed “12 March 2004, as adopted”, and it differs from the text the Economic and Social Council endorsed and published as agreed conclusions 2004/11. Most differences between the two are house style. This one is not: the draft says some men and boys themselves face barriers, and the endorsed text says men and boys sometimes face them. The endorsed wording is the weaker of the two for this campaign's argument, and it is the authoritative one, so it is what we now quote. Caught by a reader screenshot of page 147 of the endorsed text.

  The register, the research agenda, llms-full.txt, the submission document and the draft article, English and Norwegian

2026-08-15 · 7 corrections

• wordingThe newsletter box said "One click to leave, any time."

  Now: "You get a one-click removal link the moment you sign up, and it is in every mailing."

  The page promised one-click unsubscribe when the site's own API had no unsubscribe route at all. The route was built and tested first, and the copy now describes what signing up actually does.

  Newsletter box on the front page, English and Norwegian

• wordingForty to one, in one of the world's most equal countries.

  Now: The clause is gone. The page states the two figures, their two sources and the calendar year, and nothing more.

  No source was carried for the superlative and none was found. FACTS.md now bans adding it back.

  The quietest number (suicide.html) and data/suicide.json

• sourceThe Norway 40 to 1 comparison cited a single Kripos homicide report for both of its numbers.

  Now: The 482 male suicides cite FHI's Dodsarsaksregisteret. Kripos is cited only for the homicide count.

  The Kripos report contains none of the suicide figures. FHI publishes 658 suicides for 2021, and the 482 male and 176 female split in FHI's statistics bank reconciles to that total.

  The quietest number (suicide.html), data/suicide.json and the register (sources.html)

• sourceTen source quotes on the suicide dossier were document titles or descriptions, printed inside quotation marks as if the source had written them. One read "Marital dissolution and suicide risk."

  Now: Seven carry a verbatim sentence from the source, for example "Divorced men were 2.82 times more likely to die by suicide than married men." Three had no supporting sentence and are now empty, and all three page builders skip an empty quote block.

  A title is not something a source said, and inside quotation marks it reads as a quotation. The replacements were fetched from the WHO, UNODC, FHI and Wilson et al. documents themselves.

  The quietest number (suicide.html), built from data/suicide.json

• figureThe risk of suicide peaks in the first year after a separation.

  Now: Risk is greatest in the immediate aftermath. The sharpest timing figure is 13.20 times the likelihood in the 30 days after a separation is lodged.

  Wilson et al. 2025 in Psychological Bulletin names no first-year peak and reports four studies where the raised risk fades over time. The 30-day figure is Naess et al. 2021, cited inside that review. The correction reached the dossier first and the front page a few hours later, during which the site stated both versions.

  The quietest number, the front page, the story beats and the register, both languages

• figureDivide the seconds in a year by that number and the interval is 64.0 seconds.

  Now: A 365 day year holds 31,536,000 seconds, so the interval is 63.9086 seconds, which the page rounds to 64 in prose.

  31,536,000 divided by WHO's 493,456 male suicide deaths for 2021 is 63.9086, and only a 365.25 day year gets near 64.0. The live tally now counts on the exact interval, and "one every 64.0 seconds" is banned in FACTS.md.

  The quietest number (suicide.html) and data/suicide.json

• figureTwelve victims of partner or family homicide, out of 29 homicides in Norway in 2021.

  Now: Kripos registered 23 homicide cases with 29 victims in 2021, of whom 12 were killed by a partner, an ex-partner or a family member.

  Kripos, Nasjonal drapsoversikt 2021, counts cases and victims separately. Reading 29 victims as 29 homicides misstates the source.

  The quietest number (suicide.html), data/suicide.json and the register (sources.html)

2026-08-14 · 33 corrections

• wordingMovember has raised over a billion dollars for men's health, largely from men.

  Now: Movember has funded more than a billion US dollars of men's health work.

  Movember's own About us page reports funds committed to programmes, more than AUD 1.8bn since 2003, which is not the same as money raised. No source supported "largely from men", so it was dropped.

  Front page objection card on men organising, its receipts, and the register, both languages

• wordingThirteen UN days name women or girls in their titles.

  Now: Thirteen UN days are dedicated to women or girls.

  Checked against the UN's live list of international days, not all thirteen carry the words in the title, including the day to end obstetric fistula and Widows' Day. The claim is about which group a day is for, not what it is called.

  Front page days section, both languages

• wordingA badge on the newest UN days read "added in the past year".

  Now: It reads "one of the two newest days".

  The badge sat on days added in 2024 and 2025 per the UN list, so read in August 2026 it was already wrong for the 2024 day and would keep decaying. The replacement does not depend on when it is read.

  Front page days section, both languages

• wordingNorway wrote men into shelter law in 2010. Denmark in 2024.

  Now: Norway's shelter law has covered men since 2010. Denmark followed in 2024.

  Krisesenterlova dates from 2009 and its first section requires provision for women, men and children, with the municipal duty in force from 2010. Men were not added to an existing law.

  Story beat on the front page and the film (watch.html), both languages

• sourceTrue. The General Assembly and the diplomatic corps have always been majority male. No source given.

  Now: True. The diplomatic corps has always been overwhelmingly male. Only 7% of the UN's permanent representatives since 1947 have been women.

  The original carried no source at all. The figure now comes from GWL Voices' count of more than 2,800 permanent representatives, cited in the register.

  Front page objection card on who runs the UN, and the register

• sourceMale and female suicide rates of 12.4 and 5.9 per 100,000, linked to WHO's age-standardised indicator page.

  Now: The same figures, labelled crude rates and linked to WHO's crude-rate indicator page. The Norwegian register row still carries the old label and link.

  12.4 and 5.9 are crude rates in the WHO Global Health Observatory, where the age-standardised pair is 12.3 and 5.6. The row was pointing readers at an indicator that does not produce the numbers printed beside it.

  The register (sources.html and js/i18n.js). Corrected in English only

• sourceOne register row carried the European Women's Lobby's membership count and its EU funding, both cited to the lobby's own About us page.

  Now: Two rows. The membership count keeps the About us page. The funding row cites the 2026 operating grant under the CERV programme (CERV 2025-OG-SGA, DG Justice via EACEA).

  The About us page does not evidence the funding half of the claim. A claim about EU money needs the EU funding record behind it.

  The register, both languages

• sourceThe US deaths-of-despair figure was credited to AIBM, Richard Reeves.

  Now: Credited to AIBM, Rachel Kleinfeld. The Norwegian register row still credits Reeves.

  The AIBM commentary at the linked URL is written by Rachel Kleinfeld. Reeves founded the institute but is not the author of that piece.

  The register (sources.html and js/i18n.js). Corrected in English only

• sourcePizzey later described receiving threats and pickets, sourced to a Guardian link and to her own accounts.

  Now: Death threats, police protection on the book tour, and years spent outside Britain, sourced to The Independent, 1997.

  The Guardian URL was dead and her own accounts are not something a reader can check. The Independent's 1997 interview carries the specifics.

  Front page history section and the register, both languages

• sourceA 2012 lecture by Warren Farrell in Toronto needed police protection after protesters blocked the doors, sourced to contemporary reporting including the National Post.

  Now: On 16 November 2012 protesters barricaded the doors, police formed a cordon so the lecture could proceed, and one protester was arrested. Sourced to The Varsity and Maclean's.

  The Varsity's report of 17 November 2012 records the barricade, the cordon and the arrest. The earlier version was vaguer than the record and rested on a second-hand source.

  Front page history section and the register, both languages

• sourceAfter an open letter of protest from staff and students, the University of York withdraws its International Men's Day plan. The BBC link did not resolve to the story.

  Now: After an open protest letter with about 200 signatures the university apologises and withdraws the plan, and a counter-petition of more than 1,000 asks it to reconsider.

  The BBC's York report carries the letter, the apology and the larger counter-petition, which cuts against the campaign's own framing. The dead link was replaced with the working one at the same time.

  Front page history section, the register and the pushback dossier (resistance.html)

• sourceInternational Men's Day is observed in more than 90 countries.

  Now: Observed around the world since 1999.

  No source could be found for the 90-country count. The number was dropped rather than kept unsourced.

  Front page day-search prompt and the day detail card, both languages

• scopeA European Men's Lobby does not exist, in EU registers or anywhere else.

  Now: It does not exist in EU registers, and nothing comparable receives EU core funding.

  The second half was a universal negative that no register can verify. What is checkable is the EU transparency register and the CERV core funding record, so the claim was cut back to those.

  Front page institution ledger, both languages

• scopeNo equivalent instrument exists for the 81% of homicide victims who are male.

  Now: No equivalent instrument exists for the 81% of the world's homicide victims who are male.

  The 81% is UNODC's global figure in the Global Study on Homicide 2023. It sat beside EU figures on the treaty card and in the press numbers, where it could be read as European.

  Front page Istanbul Convention card and the press numbers on sources.html, both languages

• scope1 in 3 women worldwide experience physical or sexual violence in their lifetime. Source: WHO, 2021.

  Now: 1 in 3 women worldwide experience physical or sexual partner violence, or sexual violence by others, in their lifetime. Source: WHO, 2023 estimates.

  The WHO estimate covers partner physical or sexual violence plus non-partner sexual violence, not all violence, and the current release is the 2023 one. The register row now carries 31.6% and about 840 million.

  Front page context tiles and the register, both languages

• scopeNorway adopted the 35 measures its Men's Equality Commission proposed.

  Now: Norway's commission proposed 35 measures. The government has not adopted them.

  NOU 2024:8 is the Mannsutvalget's report to government, delivered on 24 April 2024 (regjeringen.no), and a commission proposing is not a state adopting. The old form is now printed on the site's own refuse list and banned in FACTS.md.

  Front page objection card on budgets, the FAQ structured data and the film (watch.html), both languages

• scopeAIBM funded with a $20M fund in 2024.

  Now: Melinda French Gates gave AIBM's founder Richard Reeves a $20M fund (2024).

  The Pivotal announcement shows the fund went to Reeves as one of twelve leaders each given $20 million to direct, with reporting putting about $5M with AIBM. "$20M to AIBM" is banned in FACTS.md.

  Front page objection card receipts and the register, both languages

• scopeEngland: 33 refuge beds for men. 4,551 for women.

  Now: England, May 2024: 33 refuge bedspaces dedicated to men, against 4,551 for women.

  The Women's Aid Annual Audit 2025 counts bedspaces dedicated to men in the Routes to Support census, as at May 2024. A further 374 admit either sex, so "beds for men" understated what men can access.

  Front page shelter card, the story beats, the film (watch.html) and the register, both languages

• scopeOfficial surveys count men as roughly three or four of every ten victims of domestic abuse.

  Now: In England and Wales, the official crime survey counts men as roughly three or four of every ten victims of domestic abuse.

  The share comes from the Crime Survey for England and Wales (ONS, year to March 2025). Unscoped, it read as a claim about official surveys everywhere.

  Front page shelter card and the story beats, both languages

• scopeFor every 100 degrees women earn, men earn 71.

  Now: In the US, for every 100 bachelor's degrees women earn, men earn 71.

  NCES Digest table 318.10 counts US bachelor's degrees for 2021 to 2022: 835,316 to men against 1,179,719 to women. Two checks the same day added the country, then the degree level.

  Story beat on the front page, both languages

• scopeThe 93.5% of workplace-death victims and three in four deaths of despair are not CEOs.

  Now: The 93.5% of EU workplace-death victims and three in four US deaths of despair are not CEOs.

  The workplace figure is Eurostat hsw_mi08 for the EU in 2023, and the deaths of despair figure is American, from AIBM. Neither is global, and the sentence read as if both were.

  Front page objection card on male power and the FAQ structured data, both languages

• scopeWomen are about 21% of UN permanent representatives today.

  Now: About 21% in 2024.

  The GWL Voices report gives a 2024 figure. The word today claimed a currency the source does not.

  Front page objection card receipts and the register, both languages

• scopeEngland's men's health strategy came with 3.6 million pounds for suicide prevention among middle-aged men.

  Now: 3.6 million pounds over three years.

  The strategy published on 19 November 2025 (GOV.UK) states the sum as a three-year total. Without the period it reads as an annual figure.

  Front page precedents timeline and the register, both languages

• scopeErin Pizzey opens the world's first women's refuge, in Chiswick, London.

  Now: Erin Pizzey founds Chiswick Women's Aid in west London. Britain's first women's refuge, widely described as the first of its kind anywhere.

  BBC News (2021) documents Britain's first. The global superlative is a description reporting repeats rather than a fact any source establishes, so the page attributes it instead of asserting it.

  Front page history section and the register, both languages

• scopeWarren Farrell, once a board member of the National Organization for Women.

  Now: Warren Farrell, who once served on the board of NOW's New York City chapter.

  The Baltimore Sun (1993) records three terms on the New York City chapter board in the early 1970s, not the national board.

  Front page history section and the register, both languages

• scopeA slider compared donor spending to EIGE's annual budget as a plain multiple.

  Now: EIGE's annual budget is named with its currency: EUR 9.4M, currencies as stated.

  The slider sets US dollar aid figures against a euro budget taken from EIGE's final annual accounts 2024. The caveat says so rather than implying a converted comparison.

  Front page money section, both languages

• scope700+ women die every day from preventable causes linked to pregnancy and childbirth.

  Now: 700+ women died every day in 2023.

  The WHO maternal mortality fact sheet the row cites gives a 2023 figure, about 260,000 in the year. The undated present tense claimed a currency the source does not.

  Front page context tiles and the register

• scopeDrafted but never published: family courts are biased against fathers, and the Strasbourg case law proves it.

  Now: Never published. The site claims enforcement failure only, and prints the refusal on the front page.

  The campaign's own red team banned it before publication. Every judgment the site cites rests on a state failing to enforce contact under Article 8, and in Terna v. Italy the Court rejected the discrimination complaint while still finding the enforcement violation.

  Front page refuse list, the family law dossier (familylaw.html) and FACTS.md

• figureAcross 33 studies in 30 countries, separated and divorced men face nearly three times the suicide odds of married men.

  Now: Across 75 studies in 30 countries, divorced men face nearly three times the odds. Separated men, nearly five times.

  Wilson et al. 2025 in Psychological Bulletin covers 75 studies, 29 of them meta-analysed, and reports 2.82 for divorced men and 4.82 for separated men. The single merged number undercounted the review and hid the difference between two groups.

  Front page suicide section, the film (watch.html) and the register, both languages

• figure6:1 or more. For every person killed by partner or family violence, six men died by suicide.

  Now: About 5.7 to 1. For every person killed by partner or family violence, nearly six men died by suicide.

  WHO's 493,456 male suicide deaths for 2021 against roughly 85,000 partner and family homicide deaths in the UNODC and UN Women femicide brief gives 5.7, not 6. Rounding up rounded in the campaign's own favour, and 6:1 is banned in FACTS.md.

  Front page suicide section and the story beats, both languages

• figureAustralia: second country in.

  Now: Australia: third country in, after Brazil in 2009.

  Brazil's PNAISH men's health policy was created by Portaria no. 1.944 of 27 August 2009, before Australia's 2010 policy. Ireland 2008 remains first, and the register row now names Brazil too.

  Front page precedents timeline and the register, both languages

• figureEIGE. Established 2010.

  Now: EIGE. Established 2006.

  The institute was created by Regulation (EC) No 1922/2006, confirmed by its own final annual accounts. The ledger card had carried 2010.

  Front page institution ledger, both languages

• figureTwo of the thirteen UN days were added in 2025 and 2026 alone.

  Now: Two of the thirteen were added in 2024 and 2025 alone.

  The resolutions date the additions. A/RES/78/323 was adopted on 13 August 2024 and A/RES/80/10 on 26 November 2025, per the UN list of international days.

  Front page days section, the callout and the day detail card, both languages