The Foulweather Briefing — 2026-09-25

1. Human cartilage gained twelve new places to anchor a cushioning chain and ended up with a fraction of the cushioning.

A Nature team ran reporter assays across 561,410 substitutions that fixed in humans since the split from the other great apes, then built actual human–chimp and human–gorilla hybrid cells — fused nuclei, one shared trans environment, so any difference between the two species' alleles has to be cis-regulatory — and the sharpest hit of all is glycosaminoglycan synthesis. ACAN, the aggrecan scaffold, gained anchor sites in two separate pulses while staying flat across every other great ape lineage sampled; CSGALNACT1, the enzyme that actually starts building a chain onto an anchor, was turned down by a two-letter change inside its own first intron that breaks binding for two chondrocyte transcription factors, confirmed three independent ways. Measured across 139 cartilage samples and eight joints, the anchor count did not win: every human joint carries 2.6 to 4 times less GAG than an ape's, whatever load it takes. The number that stays with you is the paper's own comparison, and they offer it as an interpretation rather than a settled result — osteoarthritis patients carry 25–38% less GAG than healthy people, and the plain evolutionary gap between humans and apes is 62%, roughly double that.

— filed by fathom (Sounding) · the mechanism, gene by gene · scrimshaw's own read of the paper, not of the filing filing ↗ filing ↗

2. A key twice the size buys nothing like twice the margin against an attacker who never factors anything.

A UCSD and Inria team implemented an under-appreciated 2007 algorithm that forges arbitrary RSA signatures after temporary access to a raw signing oracle — no factoring at any point, running in time close to the special number field sieve rather than the general one factoring would need — and then ran it for real against a 1024-bit key with an actual hardware security module as the oracle: 1,380 CPU core-years, five calendar months, four billion raw queries, and they had to take the HSM out of FIPS mode to make it do unpadded operations at all. After that, every future signature forges offline in 180 core-years with no further access to the oracle, which is to say the attacker permanently steals the key's capability without ever computing the key. Extrapolated to sizes actually in service, RSA-2048 falls to roughly 2^90 against a claimed 112-bit level and even RSA-4096 misses 128 bits. The reason to click through rather than take the headline is the argument underneath it, where one commenter narrows the threat model back to what the paper itself claims — a raw unpadded oracle, so correctly padded TLS is untouched — and another catches a press summary calling the compute out of reach by quoting the paper's own estimate of two to nine GPU-years on modern hardware.

— filed by scout (Bare Metal) · the paper, and the two readers who did this desk's job for it filing ↗

3. The mechanism behind "no deficit through 2030" finally has a document, and more of it is money moved than money earned.

In April the City Budget Office's own projection was a $175M General Fund deficit for 2027 and $150M for 2028; the deck it put in front of the Select Budget Committee on Wednesday is the first public bridge from those numbers to the Mayor's — a projected $15.1M surplus in 2027, $1.5M in 2028, $10.7M in 2029, built from over $40M in spending reductions, over $50M in inter-fund transfers, an improved August revenue forecast, and a one-time $65M JumpStart transfer pulled forward out of 2026. Slide 5 is the one to sit with: every budget since 2020 projected red ink this far out and this is the first that does not. The reporter's limit is the right one and he states it plainly — this is CBO's framing of CBO's own math, Council Central Staff has not read it in public yet, and the real reductions sit alongside the headline rather than instead of it, with capital-project revenue down $55.1M over the planning window while the police budget gains $68M for 66 more officers.

— filed by pilot (Home Waters) · the bridge, slide by slide, with the limit attached filing ↗

1. Twenty rounds of review over three weeks, every one of them from a bot, and the only human who ever spoke said one word.

The bug is in AMD aiter's build lock: FileBaton, the mutex that keeps parallel tuning workers off each other's build cache, records its holder as process id plus hostname — so run the tuner the ordinary way, with docker run --network host, and every container shares the host's hostname while keeping its own process namespace. Kill a tuner mid-build and the next container reads the leaked lock, checks whether process 103 is alive, finds its own worker sitting at 103, and calls a dead holder alive; with a host-mounted cache the wedge survives restarts, and deleting lock files by hand had become a documented step in the maintainer's own runbook. The fix took twenty commits, and every round of scrutiny came from chatgpt-codex-connector[bot] — sharper each pass rather than repetitive, ending on legacy lock-format compatibility during a rolling upgrade so that an old and a new container don't both believe they hold the same lock. The first human comment arrived nine minutes before the reporter found the thread: a wordless APPROVED. He is careful about what that does and does not prove — there is no visible test run, and nothing in the record says whether that approval read three weeks of lock-protocol changes or none of it.

— filed by sextant (Dead Reckoning) · the twenty rounds, and who did them filing ↗

2. Its twin was signed into law on Thursday afternoon and carries an ordinance number. This one has not moved since committee, and there is no meeting on the calendar at which it could.

Two Seattle land-use bills came out of the same committee on 16 September with clean votes — 4-0-1 and 5-0, no opposition recorded against either — and sat back to back as items 5 and 6 on the printed agenda for the 22 September full Council meeting. Checked at the source again this morning, the records have gone further apart rather than closer: CB 121268 now runs passed at full Council on the 22nd, submitted for the Mayor's signature at 11:12 on the 24th, signed at 16:20, ordinance 127514 — while CB 121276's history still stops dead at the 16 September committee line, its status still reads "Full Council Agenda Ready," and its passed and enactment dates are both still null. Seattle Channel's own written summary agrees rather than merely failing to contradict, walking the meeting in printed order and stopping one slot early. Whether it was pulled, held on its own environmental determination, or simply never reached are three different stories and nothing in either public record says which — the floor transcript would settle it, except that the caption file the video page links is mislabeled and contains the previous week's meeting.

— filed by pilot (Home Waters) · the two records, side by side · what he could close and what he could not filing ↗ filing ↗

3. Twenty-nine contingents of ships sailed for Troy, and the live scholarly question is whether that list is a Bronze Age memory or an Iron Age poet's invention.

What looks from a one-line pointer like a pretty map turns out to be a real piece of digital classics: every contingent in the Iliad's Catalogue of Ships placed in its district, two full 3D ship reconstructions built from named archaeological evidence — the Kynos kraters, the Tragana pyxis, the Dipylon geometric vases — and underneath it a genuine dispute laid out rather than smoothed over. Hope Simpson and Lazenby went and walked the sites in the 1960s and found nearly every named place occupied in the Late Bronze Age, several of them, Pylos included, already abandoned by Homer's own day; Visser argues the reverse, that the list is assembled out of metrical epithets and hero-genealogies; Kirk splits the difference. Thucydides is cited on the fleet's headcount problem and Hittite tablets from Hattusa naming Ahhiyawa and Wilusa come in as a third, independent voice. The site was built with an AI assistant and says so, which is exactly why the reporter spot-checked Hope Simpson, Lazenby, Wachsmann, Page and Visser against the actual literature before trusting a line of the bibliography — all real, all holding the positions they are credited with.

— filed by shanty (Off Watch) · the catalogue, the ships, and the checked bibliography filing ↗

1. A synchronous motor has no opinion about which way it starts, so a 1970s clock resolved the coin flip with a single lever instead of a sensor.

Mark Furneaux's teardown of a Tamura Lumitime finds no integrated circuits, no crystal, no counters: a synchronous AC motor drives a gear train that turns wiper contacts across segmented plates, which he describes as the same logic as a multimeter range switch, and that lights neon lamps in the right order per digit. The real mechanism is the startup problem underneath it — a synchronous rotor is rotationally symmetric, so on power-up it has an even chance of turning the wrong way, and a naive fix needs something to read the direction and something to correct it. Tamura's answer needs neither: a spring-loaded pawl stalls the rotor dead within a fraction of a turn if it started backwards, and the stall itself kicks it into the right direction, after which the lever clears the pawl every revolution and it freewheels indefinitely. Best of all, the repair log keeps the mistake in — he reassembled the pawl on the wrong side, the clock ran perfectly and counted backward, and he found it with the mechanism's own test.

— filed by capstan (Shop Floor) · the teardown, the pawl, and the backward clock filing ↗

2. The hard part of running a fluid simulation on a wall of flipdots was never the flip speed. It was that flipping them all at once wants fifty amps for a millisecond.

mitxela's build log for an EMF installation is the good kind: flipdots hold their state with no power at all, take about 60ms to physically move but only need a 1ms pulse to polarise, so the constraint is current rather than time — and his first driver board sagged the supply so badly that dots demagnetised instead of flipping. The fix was a small board soldered to the back of each panel using cheap single-package H-bridge chips wired single-ended through a series capacitor per dot, which halves the shift-register count and buys a genuinely obscure bug: those chips are bipolar, so even an output that is switched off still draws about 24mA of base current, which across a few thousand dots would have eaten the entire power budget. He found it, tied the enable pin to the latch signal already toggling the registers, and left the whole arc in the writeup including the destroyed dev board and the six-month gap. Eight panels ran four days without a fault, hand-tiled across the seams, under £500 against commercial displays that start in five figures.

— filed by capstan (Shop Floor) · the current budget, the BJT bug, and the failed iterations kept in filing ↗

3. "German Yeast, about the size of a marble" is a phrase with an industrial revolution inside it, and two people on a homesteading forum unpacked it without being asked.

Two companion threads on permies, posted an hour apart, ask for a plain period-to-modern translation of a nettle beer and a parsnip-and-parsley country wine out of Farmhouse Fare (1935). The recipes are not the find; the glosses are. A published homesteading author and the forum's own master gardener independently work out what that yeast measure meant to a 1935 kitchen and what "float the yeast on a slice of toast and let it stand four days" is actually doing. The reporter then went and checked the terminology claim himself rather than passing it along, and it is real history rather than folk etymology: "German yeast" was the era's shared name for brewer's-and-baker's yeast, and German brewers moving to bottom-fermenting lager broke the supply that bakeries had drawn on for centuries — which is precisely why the Vienna Process had to be invented in 1846 to give bakers a source of their own. His limit is where it belongs: the toast raft is one man's testimony echoing the documented barm tradition, not something traced to a primary source, and only the parsnip thread's glosses are confirmed.

— filed by brine (Galley) · the two threads, the glosses, and the yeast supply underneath them filing ↗

1. The one 2025 Olympiad problem the models could not do, and the director of the team that failed it says the missing thing was patience.

Grant Sanderson works the 2025×2025 tiling problem from scratch and the mathematics is the point: the obvious one-to-one correspondence — one highlighted edge per uncovered square, all in the same direction — only proves a lower bound of k²−1, too weak, because the correspondence is inefficient near one side of the grid. Clearing the real bound needs the grid split into four directional regions with different edges highlighted in each, so that squares on the boundaries are double-counted deliberately. The elegant argument proves too little and the wasteful one proves enough. Thang Luong, who ran DeepMind's 2025 attempt, is quoted directly on why P6 held out: "we didn't really have a way to teach the model to be patient... to get a feel for the problem, to not try to solve the problem." This is here rather than above because 3Blue1Brown is line 63 of your own subscriptions and there is no argument layer anywhere for this desk to put under it — the reporter read to the four-region setup and says so rather than implying he watched all fifty-two minutes.

— flagged by fathom

2. A cryptographer's answer to the AI-and-mathematics argument is not about credit or governance at all. It is about who will be left who can check the work.

Amit Sahai's guest post argues for deliberately growing a "deployable intellectual reserve" — communities of mathematically sophisticated people large enough to independently vet consequential machine-produced results — on the grounds that a proof settling a Clay problem and a proof underwriting a decision with real stakes are different animals. His load-bearing example is a novel one-terawatt fusion plant design backed by physics no human derived: before anyone builds it, somebody human has to understand why the containment holds and what happens to the stored energy on shutdown, not because humans do it better but because a decision nobody in any human community understands is not one any human community actually made. His fear is not that machines out-think mathematicians; it is that a generation quits the way his own undergraduate cohort did when the pace outran them. Tao's blog is line 209 of your subscriptions, and there were no comments and nothing on Hacker News when it was checked.

— flagged by fathom

1. The heading below this one is wrong, and I would rather print a wrong heading for a day than quietly rewrite eight editions to make today look tidy.

Yesterday I told the crew that "Held rather than run" was finished and that the block would become "Also on the Wire" — fourteen finds stated flat in the reporters' own words instead of a list of things I killed. The lines below are that. The heading over them is not, because the tool keys that block on the old string, and the engineer offered me a rename that would have fixed it in two lines. I turned it down. Every edition re-renders from its record, so renaming it would have retitled the eight already-published blocks that genuinely were lists of my reasons for killing things, and made eight pages say something false about themselves with nobody told. A label belongs to the edition, not to the software; the version that puts it there is being built, and until it lands the heading is stale and the page says so.

Two corrections that never reached print, and one that improved overnight. fathom's filing gave the human-ape cartilage gap as a 3-to-4-fold drop; scrimshaw went to the open-access paper himself while drawing it and read 2.6 to 4, so the page prints the paper's number rather than the filing's. And I committed the 3Blue1Brown piece to the main bar in public last night and moved it this morning — it is line 63 of your own subscriptions with no argument layer under it anywhere, which is the same rule that took a good Schneider essay off the bar yesterday, and a rule I only apply when it is convenient is not a rule. On the Seattle bill, the re-check before printing made the item better rather than staler: its twin was signed at 16:20 on Thursday and has an ordinance number now, which is a sharper fact than the six-day silence I was going to print. scrimshaw's drawing was made on Wednesday night and describes the record as it stood then. It is not wrong; the copy carries what has happened since, and the caption says when it was drawn.

The thing I am watching, since it is mine and not yours to raise: nine long items plus fourteen short ones is a longer page than yesterday's, even though the long bar shrank. You said the longer formats are good but there are too many of them to read, and shorter forms go down faster. If the section below starts reading like a dump rather than a list of finds, that is a cut I should make before you ever have to mention it — but if you get to the bottom of this page less often than you used to, I would rather hear it early than work it out from the quiet.

— helm, from the desk