Stimpunks Foundation × More Realms · Changelog Archive

October 2026

1 dated entry from October 2026 — pieces added, pieces revised, and every fact-check, corrections included. One month of the changelog.


This log is backfilled from the collection's full commit history and kept up from here. It records three kinds of change: pieces added, pieces substantially revised, and fact-check and attribution audits — including the errors we found in our own work and exactly how we fixed them. Small typo passes and styling tweaks are left out; anything that changes what a piece claims is in.

Why publish our corrections

We braid real science with ideas credited to named thinkers, so the facts and the attributions have to be right. We're human, and we've gotten things wrong — a paraphrase dressed as a Sagan quote, a Martin Luther King Jr. line credited to Baldwin, a forest-wide fungal network asserted as settled fact because it rhymed so well with mutual aid. Naming those in public is part of the method, not an embarrassment to bury.

The working guidelines and the full per-piece ledger live in FACTCHECK.md. If you spot an error, open an issue or reach us at stimpunks.org. Corrections are mutual aid.

2026 · October 30

Glimmer Wire, edition ten: the archive that carried every full text here began screening its callers

The weekly scan behind Glimmers, at its own page. Eight filed, six held, three seeds. Median lag 58 days, range 1 to 169 — the widest spread this page has recorded. The first item in Held is the story most squarely ours — a paper on how Autistic brains build face representations, held because the publisher answers with an authentication redirect and the only readable account describes what we do as what we fail to do.

SitePubMed Central now answers article paths with a bot check, and we did not try to get round it

Five consecutive editions read every one of their full texts at PubMed Central. This week a plain request for an article path returned, at HTTP 200, a Google reCAPTCHA interstitial titled Checking your browser rather than the article — on all six articles tried and on the ?report=classic variant. We did not attempt to defeat the challenge and no edition will. A bot check is a publisher saying who may read automatically, and getting round one would be a worse thing to publish than an unread paper. The archive is screening rather than closed: its idconv identifier service answered normally in the same minute, resolving thirteen DOIs in one request, and the same six articles opened in full through an ordinary page-fetching tool. Five versions of record were read end to end that way. The cost is stated at the top of the page rather than in a footnote: last week any reader could repeat our reads with one command, and this week they cannot, and a method a reader cannot repeat is weaker evidence than one they can.

NewEight filed — four verified, three plausible, one contested

Graded verified, all read end to end: Chiarenza et al. in Science Advances, who project niche models across 245 million years and 402 tips and find that crocodile-line archosaurs came through the end-Cretaceous because their requirements had scarcely moved and a place that met them persisted — “unlike non-avian dinosaurs, low-latitude suitability persisted across all but the most extreme end-Cretaceous scenarios” — while refusing the tidy version of their own result in the abstract’s last line: “ancient in crocodile-line archosaurs, but not completely static”; Wang et al. in Nature Photonics, who push X-ray high-harmonic emission past a ceiling that has stood “for over three decades”, observing “a weak secondary plateau that extends the conventional cutoff beyond 120 eV up to the water window at 280 eV”; Postat et al. in Nature Immunology, who show activated T cells reading the stiffness of the tissue around them and changing what they become, “modulating the expression of the tissue-resident memory T cell-associated transcription factors Klf2, Runx3 and Hic1” — while stating plainly that “the mechanosensor(s) responsible remains to be determined”; and Dodds et al. in Science Advances, who re-examine seventy years of valence–arousal–dominance and conclude that “the power-danger-structure framework is the minimal framework that represents essential meaning”, two of the original three axes being dispensable.

Three are plausible on complete deposited abstracts, because none of those publishers would serve the paper. Geiger et al. in PNAS demonstrate circadian rhythm, sleep architecture, REM-like bouts, raised arousal thresholds and rebound after disruption in the spider Portia fimbriata — opening with the observation that “most of what we know about sleep is heavily vertebrate-biased” — and conclude only that “the behavior we observe in spiders is behavioral sleep”, where the wire headline asks whether spiders dream. Kojima et al. in Current Biology film dwarf reed snakes rolling and find they “stabilize their head during dynamic rotational motion of the posterior body by using precisely coordinated body movements”, which the authors read as “behavioral innovation” circumventing “deep morphological constraints”. Burriel-Carranza et al. in Royal Society Open Science report the first biofluorescence in the fire salamander, with the glandular fluorophores “also ubiquitous in blood components…a feature previously known only in tree frogs”.

One is contested, and the grade is about a verb rather than the work. Schrader et al. in Science Advances analyse “163 genomes spanning 12 of 16 extant ant subfamilies” and report convergent transposable-element bursts preceding the post-K–Pg radiations. We read the paper to settle the grade and its own Discussion settled it against its abstract: “All evolutionary genomic studies are inherently limited in their ability to establish causal relationships”, and the extinction “could have independently promoted both lineage diversification and TE proliferation” — a confounder the authors name and cannot exclude, under an abstract that says the findings “demonstrate the broad-scale impact”.

Fact-checkA lag of 169 days that is really 1,837, and six held

The longest bar on the page is measuring the wrong event. The ousiometrics paper was posted to arXiv on 13 October 2021 with an abstract identical, sentence for sentence, to the one the journal published on 8 May 2026. The wire reached it on 24 October 2026, so the lag reads 169 days by journal publication and 1,837 days by when the finding became readable — against a previous record on this axis of 753. The page says so in its own hero and in the item.

Six held. First, and the reason the section exists this week: Griffin et al. in Nature Mental Health, decoding whole-scalp EEG in roughly 400 Autistic children, held because nature.com answers the article path with HTTP 303 to an authentication endpoint, no abstract is deposited at Crossref or held by Semantic Scholar, there is no PubMed Central copy and no preprint we could find — no sample size for the comparison group, no effect size, no statistic and no stated limitation exists outside a subscription — and because the only available account frames the result as responses “less clearly distinguishable” in Autistic participants who “did not show the same pattern”, a difference stated solely as a shortfall. We will not grade a claim about our own neurology from a press release, and we will not repair it into a gift either. The other five: a cold-atom experiment testing relational constructions of time, whose abstract claims only “a controlled experimental setting in which relational-time constructions can be quantitatively tested”, carried as physicists creating a tiny universe; a Friedmann-instability theorem in Proceedings of the Royal Society A offered as an observational alternative to dark energy, with no data confrontation in its abstract and 155 days old; a post-hoc polygenic subgroup inside the ASPREE trial reported as a 70% lower dementia risk, where the absolute difference is 2.3 percentage points and major bleeding rises further in the same stratum, under a paper whose own title says may modify; a Mycologia phylogenomics paper whose title says it “suggests a novel psychoactive lineage”, carried as a mushroom that makes people see hundreds of tiny humans, with no abstract reachable anywhere; and an Astrophysical Journal Letters paper of 2 December 2025 on the 709 BCE eclipse, lag 328 days, carried in the present tense — the fourth consecutive edition to hold an item of that shape.

The picture census ran by the fixed method and caught the method moving. Nine ScienceDaily topic feeds were requested and strange_science answered 404, so eight carried content, deduplicating to 88 unique releases in the seven days to 30 October, all 88 fetched: 13 lead with an image credited AI/ScienceDaily.com and 44 more with a Shutterstock photograph, with two further releases naming an institution and then stating in the same line that the image was machine-made. The page says plainly that this is not comparable by subtraction to last week’s 56 of 91, because that sweep had one more feed in its denominator.

2026 · October 23

Glimmer Wire, edition nine: eight filed, six held, and one of our own measurements withdrawn

The weekly scan behind Glimmers, at its own page. Eight filed, six held, three seeds. Median lag 32.5 days, range 4 to 144. The first item in Held is ours — a number this page has printed for three weeks, retired because it measured how we asked rather than what we asked about.

NewEight filed — four verified, three plausible, one contested

Graded verified, all four read end to end at PubMed Central: Collins et al. in Science Advances, who restore a bleached coral by holding it against a healthy one — symbionts crossing in as few as three days by tissue fusion, full recovery within seven weeks — and then report the condition the coverage leaves out, that “contact-dependent recovery is restricted to corals of the same genet, whereas heterologous genet pairings trigger antagonistic interactions”; Johansson et al. in PLOS Biology, who find a nightjar’s broad slotted wing tip buys hovering at the price of span efficiency at cruise, and record an upstroke thrust that “challenges an existing paradigm that considers an active upstroke to primarily add weight support in birds”; Tynianskaia et al. in Science Advances, establishing the marmoset’s smooth cortex as secondary lissencephaly — a reduction from a folded ancestor rather than an unfinished version of one — and reframing the animal as “a particularly powerful system to investigate the cellular and developmental basis of secondary lissencephaly”; and Asor et al. in Nature, watching individual virus-like particles assemble and finding that “weak and reversible multivalent interactions control the assembly process”, yielding one structure “with high fidelity despite thousands of available assembly intermediates”.

Three are plausible on complete deposited abstracts, because none of those publishers would open a connection. Hollands et al. in RAS Techniques and Instruments compute the perceived colours of white dwarfs, demonstrate known stars that would look yellow, green, cyan, pink or purple “if viewed close enough for their stellar discs to be resolved”, and then report that across the 20-parsec sample “on average, white dwarfs are indeed white in colour”. The LBT Yp Project measures the primordial helium abundance at Yp = 0.2458 ± 0.0013, a precision of 0.5%, in agreement with the Big Bang nucleosynthesis value predicted from Planck’s baryon density. He et al. in Subterranean Biology describe Erhaia dijiang, the first freshwater cavesnail known from China — and its eyelessness, which the headline makes the news, is in the abstract a property of the whole family. One is contested: Fernandez, Burke & Nadel in Perspectives on Psychological Science propose that cognitive aging is “a continuation of development” rather than compensatory decline. It is graded contested because it calls itself A Hypothesis and Framework in its own title and its predictions are untested, not because anyone has attacked it.

CorrectionWe withdrew a measurement of our own that we had published three weeks running

Editions seven and eight reported the byte size of science.org’s 403 refusal week on week — 5,657 bytes, then 5,689 — and edition eight drew a conclusion from the change: “the door is the same door and the notice on it has been edited.” It does not survive a control. Requesting the identical URL twice within one minute this week returned 5,352 bytes to a bare curl and 5,643 bytes to a browser user agent, a spread nine times larger than the week-on-week difference we had treated as a signal. The body is assembled per request. We never held the method fixed, so we were never measuring the publisher. The probe is retired and no further edition will carry it; the refusal itself is still recorded, because a 403 is a real answer and a closed connection is not. The shape of the error is one this page names in other people’s work most weeks — a figure that is real, reproducible and about the instrument — and it survived three editions because it read as diligence.

A second habit broke in the opposite direction. Four consecutive editions recorded bioRxiv as unreachable on an HTTP 429 and left it there. The same 429 came back this week on the first several attempts and then, with a pause between requests and a browser user agent, two preprint pages returned 200. Four weeks of “refused” was four weeks of not retrying. The .full text paths stayed refused, so what was read is each preprint’s abstract page and the page says so.

Fact-checkSix held, and a preprint comparison that reversed what we expected to find

Reading those two preprints overturned the story we had drafted from their titles. The marmoset preprint of 19 February 2026 is titled “Coordinated neural progenitor adaptations drive primate neocortical downscaling”; the version of record says “contribute to”, which reads as peer review softening a causal claim. The two abstracts are word for word identical, and both already said “bias…toward” and “likely constrain”. Nothing was weakened: a title that overstated its own abstract was brought into line with it. We would have published the opposite from the titles alone. Preprint intervals of 245 and 261 days are recorded against journal lags of 29 and 36, which is the lag axis’s standing blind spot showing itself.

Of the five other holds: a compilation of 2,884 Type Ia supernovae reported as evidence that dark energy varies, with no DOI, no journal reference and no link to a paper on its page and no match in a Crossref bibliographic search; a Science paper of 4 November 1960 — von Foerster, Mora & Amiot’s “Doomsday: Friday, 13 November, A.D. 2026” — carried in the present tense at a lag of 24,087 days; an Entropy paper arguing that psychoanalytic projection is “closely analogous to” predictive-processing prediction, reported as neuroscience rediscovering Freud, held because the borrowing runs from the field with standing to the field gaining from it; a careful flight-initiation-distance study from 12 December 2025, lag 314 days, under a headline asserting that “scientists have no idea why” above an abstract whose closing line is that the authors discuss why; and an honest competition study between managed honeybees and wild pollinators, held on register as a harm belonging in Triggers rather than a Glimmer, with bullying noted as the wire’s word and interference competition as the paper’s.

One citation that simply goes to the wrong place. The primordial-helium story carries a single DOI, and it is Paper I of five — the project description, whose abstract states a goal of 0.5% precision. The measurement is in Paper IV and the cosmology in Paper V, both published the same day; we found them by searching Crossref for the series title. A reader following the link provided would find a plan rather than a result. Three of that release’s figures are named on the page as the release’s own, appearing in no abstract we read: “three times more precise”, “130 hours of observations”, and “15 galaxies” where Paper IV says 15 low-metallicity targets among H II regions.

The picture census continues by the method fixed last week, with a wider net. Nine ScienceDaily topic feeds deduplicated to 91 unique releases in the seven days to 23 October, all 91 fetched: 10 lead with an image credited AI/ScienceDaily.com and 46 more with a Shutterstock photograph. The page states plainly that this is not comparable by subtraction to last week’s 28 of 42, because that sweep covered eight feeds and this one nine.

2026 · October 16

Glimmer Wire, edition eight: nine papers read end to end, three of them read and then held

The weekly scan behind Glimmers, at its own page. Eight filed, eight held, three seeds. Median lag 19 days, range 0 to 236 — and the zero is literal: a dinosaur was described and reported on the same day, after waiting a hundred and two years in a museum.

NewEight filed — five verified, one contested, two plausible

Graded verified, all five read end to end at PubMed Central: Rogers et al. in Nature Communications, mapping 3D genome folding across squid, cuttlefish and octopus and coining regulatory entanglement for what they find — the accumulation of interdependencies until regions brought near each other “become embedded within nested regulatory architectures” — and then naming its cost, a hypothesis that the octopus lineage’s deeper entanglement “limit[s] the capacity for further genome expansion”; Fusar Bassini et al. in Nature, whose mass-spectrometry atlas of the mouse brain resolves 539 lipid territories that partly mirror the cell-type map and partly do not, because some take their chemistry from axon terminals belonging to cells elsewhere; Yu et al. in Science Advances, who located the residues deciding which blast-fungus effector a cereal immune receptor recognises and used them to build dual-specificity receptors that answer to effectors their parents could not see; Alcott et al. in Nature Communications, tracking phosphorus through two Great Oxidation glaciations and concluding that the feedback drove instability “persisting for over 200 million years” rather than a smooth rise; and Sanchez et al. in Nature Communications, who measured magnesium and silicon together in an ultra-hot Jupiter’s atmosphere and so gave the first empirical test of an assumption their own abstract calls “largely untested” — that a planet’s rock-forming ratios match its star’s. That one carries the longest lag this page has filed, 236 days.

Two are plausible on complete deposited abstracts, because neither publisher would open. Morrison et al. in Royal Society Open Science describe Invictibellatrix amberina, the first megaraptoran known from Europe, from a braincase the Natural History Museum acquired in 1924 — and the abstract’s methodological claim is the better half, that theropod braincases “are often regarded as morphologically and phylogenetically conservative” and are not. Yu et al. in Science record the lineage of a mouse in its own DNA, reconstructing from a single embryo “a time-calibrated, parsimony-supported, zygote-rooted phylogeny of 1.28 million transcriptionally profiled cells.” One is contested: Katlowitz et al. in Nature, recording a human hippocampus under propofol still discriminating oddball tones, improving at it over ten minutes, and carrying semantic and grammatical information about speech. The recording is solid; whether it settles anything about consciousness is what the paper’s own first sentence calls disputed, and its authors name propofol and the hippocampus as limits on their claim.

Fact-checkEight held, and three of the eight were read end to end first

The cleanest hold this page has made turns on a grammatical subject. A Nature paper titled “A Jurassic mammaliaform swimmer and transformation of the mammalian pharynx” ran as “this tiny jurassic predator helped pave the way for the human brain”. The brain appears exactly once in the coverage, in a co-author’s quoted inference three evolutionary steps long; the paper itself is closed, with no PubMed Central copy and its abstract elided by the publisher, so the only thing we could verify is that the headline’s sentence is not the title’s.

A second hold is about punctuation. A Light: Science & Applications paper we read in full ran under the headline “Physicists discover ‘astonishing’ property of light”. The word astonishing occurs six times on that page — in the <title>, in og:title, in the <h1> and three times in auto-generated citation blocks that repeat the headline — and nowhere in the body, in any quoted sentence, or in the paper. Quotation marks promise that somebody said the thing; the university’s own newsroom is unreachable from here, so we cannot say whether anyone did, and not being able to say is why it is held rather than filed or accused.

Two more of the familiar shape, and one that is ours rather than theirs. “Scientists make stunning discovery that could help explain why life exists” rests on a Science Advances paper of 25 August 2023 — a 1,141-day lag, and the third consecutive edition to hold an item carrying an old result in the present tense. “The secret to easier exercise as an adult may start in childhood” is fourteen weeks of limb loading in growing fowl, a real and striking 22% result that the abstract never extends to children, and the paper could not be read. And a National Science Review paper on the Moon’s two faces, read in full, is held because its authors write “we propose” where the headline writes “solved” — and because its sharpest sentence, that “the mantle displays no clear compositional dichotomy”, is a negative result the headline drops. Two further items — two cats with a Marfan variant, and an estimate of the Australian Indigenous population before invasion — are declined on our own grounds of register and standing rather than on any doubt about the work, and the page says so plainly.

Two claims checked against the papers and found to belong to the coverage alone. The word intelligence occurs zero times in the octopus genome paper, whose own phrase is “large and elaborately structured nervous systems, novel organs, and complex behaviours”. And habitable occurs three times in the exoplanet paper — twice inside an author’s institutional affiliation and once in a reference title — so the release’s line that the result helps investigate potentially habitable worlds has no basis in the work. A third case was filed rather than held and named on its face: the Great Oxidation release speaks of oxygen being “sustained” where the paper says instability, twice.

CorrectionWe nearly published that the DOI system was down. It was not

The reachability list on this page is built by probing each host, and the probe was wrong. Following redirects, a request to doi.org returns nothing, which reads exactly like a dead resolver — and doi.org was on its way onto the page as a failure. With redirects disabled it returns HTTP 302 and a 305-byte body carrying the correct target, every time. The resolver worked; the publisher at the end of the chain did not. A probe that follows a redirect reports the health of the last host and files it under the name of the first. The method is corrected, the correction is on the edition page in its own paragraph, and the lesson is the house’s own: our own claims are the unchecked ones.

A gate caught the other one before a reader could. The new cards on the collection page and the front page used a mono class copied from the edition page, which defines it, onto two pages that do not — an intention that would have silently not happened. check-classes.mjs failed both, which is the exact fault that check exists for. Fixed before shipping.

SiteA fourth identical byte count, a preprint the lag axis cannot see, and a second picture census

PubMed Central’s wrong-path home page measured 67,236 bytes for a fourth consecutive week, the same figure logged on 25 September, 2 October and 9 October. A genuine missing-article path returns a proper 404, so the fault is one search path rather than the archive. One of nine articles requested met reCAPTCHA at 200 in a 20,422-byte body, inside last week’s 20,416–20,485 band, and cleared on the third user agent. science.org’s 403 body measured 5,689 bytes against 5,657 last week; biorxiv.org returned 429 on every article path for a fourth week. Refused at the network layer: Royal Society, Nature, Oxford, Taylor & Francis, arXiv, Europe PMC, EBI, Unpaywall and UEA.

The lag axis’s standing blind spot showed itself, so it is reported rather than inferred. Searching Crossref for posted content under the lineage paper’s own title returns a bioRxiv preprint posted 30 July 2026, whose deposited abstract is the same text as the journal’s. The bar on the page reads 3 days; the work has been public for 73. The preprint could not be read either, so the comparison is between two abstracts and the page says so. No other filed item returned one.

The picture census was run a second time and deliberately not compared. Deduplicating eight ScienceDaily topic feeds gives 43 releases in the seven days to 16 October, of which 42 were fetched; 6 carry an image credited AI/ScienceDaily.com and 22 more a Shutterstock photograph — 28 of 42, two thirds whose picture of a finding comes from neither the research nor the institution. Last week’s 14-of-76 was built from a different sweep, and the two denominators are not subtractable; one method will be used from here. One filed item carries a machine-made image and is named in its own entry rather than quietly left out.

One open item is logged against our own pages. No. 84 opens its Great Oxidation spread with “the moment the accounting flips”. After the Alcott paper, moment is the wrong unit by two orders of magnitude. A re-reading against the paper rather than against our own summary is owed, and is deliberately not done in this pass.

2026 · October 9

Glimmer Wire, edition seven: five papers read end to end, every one of them through the same archive

The weekly scan behind Glimmers, at its own page. Six filed, six held, four seeds. Median lag 68.5 days, range 1 to 216 — and for the first time in four editions the lags are spread rather than clumped at the ends.

NewSix filed — four verified, one contested, one plausible

Graded verified, all four read end to end: Breedy et al. in ZooKeys, describing a new deep-sea octocoral family from seamounts off Isla del Coco — and reporting that three standard barcoding markers placed it in three different families, with mtMutS giving 100% ultrafast bootstrap support for a family of corals that have no axis, while phylogenomics over 153 loci and 37,654 base pairs backs a fourth reading; Horvath et al. in Nature Communications, who realised Hans Bethe’s 1931 strings — many-particle bound states that can exist only in one dimension — in caesium atoms held in thousands of one-dimensional tubes, identified by comparing an expansion inside the tubes against a release into three dimensions where the strings cannot survive; Lestina et al. in PeerJ, on a cavity-nesting wasp hollowing out the same Macaranga stems that house the trees’ mutualist ants, at 12 of 43 trees in oil palm against 1 of 41 in logged forest; and Majolo, Wakes & Ruta in Evolution Letters, who scored five kinds of aggression separately across 100 primate species and found the lethal kinds correlated with each other and with mild aggression not at all.

One is graded contested on its second claim rather than its first. Lole Durbin et al. in Science Advances show that a maximum age calibration for the origin of animals is an argument from absence, and test it by finding a second exceptionally preserved microfossil deposit — the Kheseen Biota of Mongolia, 550 to 526 million years old — that contains no animals either, at a time when animals demonstrably existed elsewhere. That part is well made. The ~800-to-700-million-year origin it supports is a clock estimate specialists will argue over for a decade, and the headline’s “200 million years earlier” appears in no abstract. One is plausible on a deposited abstract: Hobbs & Kranzusch in Science, reporting that bacterial CBASS immunity is triggered by the phage’s own prohead protease cutting an activation loop — the enzyme a phage needs to build its capsid. science.org returns 403 and there is no PubMed Central copy, so nothing but the publisher’s abstract is on the page.

Fact-checkSix held, four of them because the thing in the headline is not the thing in the paper

The one we most wanted was the one with three faults. A Communications Earth & Environment paper imaging a giant magnetofossil in three dimensions finds a single magnetic vortex optimised for sensing variation in field intensity — good evidence for an ancient navigational sense. But its Methods place the imaged particle in Palaeocene–Eocene Thermal Maximum sediments from IODP Hole U1403B, tens of millions of years younger than the 97 million years in the headline, which is the age of the oldest sediments in which such fossils are known at all; the paper attributes them to “(as yet unidentified) eukaryotes” where the coverage says animal; and its own key sentence begins “if confirmed”. Lag 345 days.

Three more holds of the same shape. A genetic-code story carried in the present tense is McGowan et al., PLOS Genetics, published 5 October 2023 — a ciliate in which UAA and UAG are reassigned to different amino acids, which is a fine result and a 1,098-day lag; the release concedes the date in its fourth paragraph and the headline does not. A Nature Communications X-ray study of proton-coupled electron transfer ran as “how nature moves energy so efficiently”; the object measured is a synthetic ruthenium polypyridyl model complex, and the release itself notes that its mechanism was already understood. And an Annals of Human Biology catalogue of 18 undescribed Drimolen fossils ran as “Homo erectus 200,000 years earlier than expected”; its own abstract says it tests existing hypotheses about H. erectus at ~2.0 Ma and that one specimen gives additional evidence for 2.04–1.95 Ma — a date published by Herries et al. in Science on 3 April 2020.

Two held for want of any source. A carcass-feeding study of zoo lions is exactly the kind of finding this site exists for — the behaviour was missing from the schedule, not from the animal — and it rests on three lions with no reachable text: sciencedirect.com refused, no abstract at Crossref, "abstract": null at Semantic Scholar, no PMCID. Its lag cannot even be stated. And a 30-year forestry result arrived as a researcher-written piece in The Conversation carrying no journal reference at all; we traced the paper ourselves from a link inside the piece and still could not read it, and the piece’s “win-win-win” frame is an argument its authors are making in public rather than a result.

SiteOne archive stood between this edition and nothing, and 14 of 76 news pictures were drawn by a machine

Every one of the five full texts read this week came from PubMed Central, and nowhere else was possible. Four of the five are CC BY — at Pensoft, PeerJ, Oxford and Nature — and all four of those publishers refused the connection outright. Last week this page wrote that an open licence is a permission rather than a road; this week the road was singular. Refused at the network layer: zookeys.pensoft.net, peerj.com, tandfonline.com, sciencedirect.com, ncbi.nlm.nih.gov, eutils.ncbi.nlm.nih.gov, Europe PMC, EBI, OpenAlex, and both nobelprize.org addresses. Refused by the publisher: science.org with a 403, and nature.com with a 303 to an authentication endpoint, in front of CC BY articles. biorxiv.org returned 429 a third week running. PMC repeated both documented behaviours — the home page at 200 for a wrong path at the identical 67,236 bytes recorded on 25 September and 2 October, and reCAPTCHA at 200 for three of seven articles, cleared by retrying with another user agent.

A measurement of our own, and it is new. Of the 76 ScienceDaily releases carried in the seven days to 9 October, 14 lead with an image credited to a machine — thirteen Credit: AI/ScienceDaily.com and one Credit: NASA; background by AI/ScienceDaily.com — and 37 more lead with a Shutterstock stock photograph. That is 51 of 76 whose first picture of a result comes from neither the research nor the institution. The credit lines are honest and that is the entire defence; the ratio is the finding, and two of the six held items carry one. Last week the thing in the abstract slot was written by a model; this week the thing in the photograph slot was drawn by one. One week is a measurement and not a trend, and the page says so.

One fact is on the page marked unconfirmed, which is the honest alternative to cutting it. We wanted to note that Hans Bethe’s 1967 Nobel was for work on how stars produce energy — the nucleosynthesis this whole site is named out of — and could not confirm the citation at its source, because both nobelprize.org addresses are refused here. It is labelled as unconfirmed on the page and should not be quoted off it.

The lead list promised a ruling and got one. It had grown for three editions with nothing coming off it. One lead was worked — the Geology paper on the Zaonega Formation, whose deposited abstract reports a thermogenic origin for a 2-billion-year-old carbon signature many have read as global, with an apparent temperature gradient from 72 to 350 °C declining upward from a gabbro sill. The other two are closed unworked and named. And no new list was started, because a list that only grows is a way of looking thorough while deferring the work.

The shallow-clone fault was present for a third week and caught before anything was generated. This checkout arrived with 50 commits of 463, dating Bone Song to 13 September 2026 instead of 17 July. git fetch --unshallow ran first; no derived file was written against the truncated history.

Wired in: a card on the collection page and on the front page, prev/next edited on both editions, a sitemap entry added, and the collection’s prose counts re-derived by summing the seven edition tallies rather than incremented — seven editions, forty-nine filed, thirty-six held, twenty-one verified.

2026 · October 2

Glimmer Wire, edition six: six papers read end to end, and the first story held because its only summary was written by a machine

The weekly scan behind Glimmers, at its own page. Eight filed, six held, four seeds. Median lag 20 days, range 4 to 91 — less than a quarter of last week’s median.

NewEight filed — five verified, one contested, two plausible

Six versions of record were read in full, which is the best this page has managed — four at PubMed Central and two at PLOS. Graded verified: Sullivan, Gerstein & Kajiura in Integrative Organismal Biology, who flew a drone over wild blacktip sharks and found them turning away from a sound source from at least 62 m, with 71.5% of 165 responses initiated in the acoustic far field — where a shark, having no swim bladder, was long inferred to be unable to hear at all; Krzewińska et al. in Science Advances, who sequenced 142 ancient genomes from three Swedish cemeteries and found that children buried with adults were almost never closely related to them, at 12.0% of multiple burials at Västerhus and 12.5% at Sigtuna; Levy et al. in Autism Research, raising the estimated prevalence of Phelan–McDermid syndrome from 2.5–10 per million births to about 1 in 7,300 — a correction whose causes are referral rates, insurance criteria and an assay that misses half the cases; Postberg et al. in Science Advances, on 961 Cassini mass spectra showing Enceladus’ ice grains sorted into at least five salt chemistries by slow freezing and wall collisions on the way out; and Seiler et al. in PLOS One, who built a papyrus scroll, wrote on it in leaded ink, carbonised it in a low-oxygen furnace and read it back by X-ray tomography, so that an algorithm for the real Herculaneum scrolls can be marked against a known answer.

One item is graded contested on the method rather than on our access. The Bat1K consortium’s new bat phylogeny — 103 genomes, all 21 families, 44 pre-Quaternary fossils — places the origin of bats and of powered flight in Europe in the late Palaeocene. A continent of origin inferred from a fossil record that is best sampled in Europe is the kind of claim specialists argue over for a decade, and the paper hedges with probably where the coverage does not. The headline’s “65 million years ago” is in neither the abstract nor the university’s release, both of which say late Palaeocene — conventionally about 59.2 to 56.0 million years ago, so the headline is roughly six million years early and puts the origin at the Cretaceous–Palaeogene boundary rather than well after it.

Two rest on publisher-deposited abstracts: a revision of the Australian stick-insect genus Anchiale that takes two species to five while resurrecting one name from synonymy, sinking another and fixing three to designated lectotypes; and satellite tracking of nine yellow-morphotype green turtles in the Galápagos, none of which left the archipelago, and whose home ranges are “highly variable among individuals” by a spread roughly twice their own mean.

Fact-checkA new kind of refusal: a fluent summary, correctly labelled, written by a language model

Where a publisher elides an abstract, Semantic Scholar’s API returns "abstract": null and, in the same object, a tldr field holding fluent, specific, paragraph-shaped prose about the paper, generated by a language model. The archive names the model, which is to its credit and is the whole defence. For the Nature paper arguing that Cretaceous zhelestid mammals are zalambdalestoids — a group read for forty years off its teeth and reassigned on a nearly complete Gobi skeleton — that text was the only summary reachable from here, Crossref deposits none and nature.com is refused. It is not a source, and nothing about its shape says so. We did not quote it; the story is held rather than filed, and the page says why. A status code is not a source, a byte count is not a page, and a well-formed field in the right place is not an abstract.

Two holds turn on a number rather than a word, and in one of them the dropped figure is the one we would have led with. A PLOS Medicine study scoring brain ageing across nine conditions reports Alzheimer’s at d = 0.97 and tobacco use disorder at 0.72 — and autism at d = 0.06 (p = 0.36) and ADHD at 0.01 (p = 0.98), which is no difference at all. Its own abstract opens “PAD was consistently greater across disorders” and reaches those two exceptions four clauses later; the headline keeps the opening. We held it anyway, because filing it would mean accepting a predicted-age-difference as a measure of brain health in order to enjoy a null, and a page does not get to switch instruments depending on the answer. And a Psychological Medicine meta-analysis ran as “cannabis users were twice as likely to commit violence”: the factor of two is its cross-sectional figure, while its longitudinal estimate — the only design that can speak to order in time — is OR 1.17.

The other three holds: a review, a date and a route. A Biomolecules paper carried as a finding about mulberry and gut bacteria is a narrative review of preclinical models, and its own verbs say so. A MIND-diet cohort result ran on 30 September having been published on 17 March — a lag of 197 days, longer than every filed item on the page — and its “2.5 years” is an equivalence computed from a grey-matter volume slope, not a measurement of brain age. And a Nature paper on 1.75-to-1.4-billion-year-old eukaryote fossils is held because the only abstract we could obtain came from a third-party index rather than the publisher, and because its mitochondria claim is explicitly an inference from where the fossils occur.

SiteWhat refused, three open-access papers we could not open, and a shallow clone caught before it shipped

Two host behaviours repeated from last week, exactly. A wrong search path at PubMed Central again returned the PMC home page at HTTP 200 — 67,236 bytes, the identical figure recorded on 25 September, for four different queries a week apart. And PMC again served Google’s reCAPTCHA at 200 for every article on the first attempt; a retry with a different user agent returned all four papers in full. export.arxiv.org was refused while arxiv.org answered, as last week.

Refused at the network layer, named on the page: nature.com, frontiersin.org, academic.oup.com, onlinelibrary.wiley.com, doi.org, export.arxiv.org, ncbi.nlm.nih.gov, Europe PMC at both addresses, OpenAlex and Unpaywall. science.org answered with a real 403 and biorxiv.org with a 429. Three of this week’s refusals sit in front of openly licensed papers — the CC BY stick-insect revision, the CC BY turtle tracking and the CC BY bat phylogeny — which is the third consecutive week an open-access paper could not be opened from here. An open licence is a permission, not a road.

The shallow-clone fault from last week was checked for before anything was generated, and it was present. This checkout arrived with 50 commits of history rather than 461, and the test recorded on 25 September — ask git log --diff-filter=A for the creation date of Bone Song — returned 2026-09-12 instead of 2026-07-17. git fetch --unshallow first, generators afterwards; the derived files were never written against the truncated history. Last week this cost a 1,463-line diff that every gate passed. This week it cost one command, because the lesson had been written down.

One fault caught before publishing, and it is the third instance of the same one. The first draft of this edition’s collection and front-page cards used <span class="mono"> for a status code — and neither the front page nor the collection page defines .mono, so the text would have rendered in the body face with nothing to show for the markup. check-classes named it on both pages. The ledger already records this against edition four’s collection card and edition five’s changelog entry; a gate catching the same mistake three times is the gate working and the author not learning, and the fix is the same both previous times: <code>, which every page styles.

A new month is a new file. This entry opens the October archive page, copied from September with its <title>, canonical, meta, og:, twitter: and JSON-LD all rewritten for the month, prev/next edited on both pages, a sitemap row added and the index rebuilt from the month pages.

Wired in: a card on the collection page and on the front page, prev/next edited on both editions, a sitemap entry, and the collection’s prose counts re-derived by summing the six edition tallies rather than incremented — six editions, forty-three filed, thirty held, seventeen verified.

Browse the collection →