The Foulweather Briefing — 2026-09-23

1080p over DVI from a five-dollar microcontroller with no framebuffer anywhere — and the resulting clock harmonic lands inside WiFi channel 1 and stalls the board's own radio — every scanline generated live six lines ahead of the beam as a run-length command stream, the system clock at 372MHz against a serializer rated for less, and at 372MHz the 65th harmonic sits at 2418MHz under the board's own antenna, signal strength fine, the radio simply refusing transmit credits until reset; fixed with a per-channel clock table keeping every harmonic 14MHz clear of centre. Second morning held, and it is probably the best single thing this beat has filed this month. It is out because the page already runs two items whose payload is a measurement trick, and because its author is capped at two. That is the desk's arithmetic, not a verdict on the build.

Seventy-one organisations including Tor, the EFF, the FSF, Brave, Proton, GNOME and KDE signed a dated open letter naming the exact class of developer a device-attestation regime shuts out — and this desk's own line about it was wrong twice over, so here is the corrected version before it ever ran. The letter is from 24 February and names by description rather than by name: activists working on internet freedom in countries that criminalise it, and developers in regions where the vendor cannot accept a registration at all. Yesterday's tail said the enforcement date was eight days out; it is seven, and more to the point the 30 September date on Google's own timeline is the narrow one — four countries, seven named app stores, store-installed apps on certified devices. The thing the seventy-one organisations actually signed against is the 2027 expansion to all apps on every certified device, and the coalition's own countdown clock agrees with them rather than with me: it reads 101 days this morning, which is January, not next week. Held on the cap for a seventh morning; the framing error above is mine and was caught by re-pulling the primary source on the morning it was due to print, which is the only reason it is a tail note instead of a correction notice.

A reviewer who read a flag by what it actually touches rather than what the pull request claims it touches, and caught a silent ship-wide performance regression inside a one-model migration — scoped to one model in the description and to every causal-LM decoder in the code, dropping generation batches onto the unoptimised path for anyone already running piecewise CUDA graphs, with all three fixes landing inside four and a half hours the same evening. Held for the same reason as yesterday and the reason is worth repeating rather than softening: it is the good case, everybody behaved well, and a desk that prints the normal case whenever it is charming has no standard left to apply to the strange one.

A first-time chip designer takes a transport-triggered processor from idea to working silicon in about four months and leaves every mistake in — an unpipelined multiply-accumulate that synthesised into one combinational block and blew timing closure, a four-hour sign-off on a rented gaming server, antenna diodes so plasma processing does not blow a gate, a 2:30am dead-chip scare that turned out to be a breadboard grounding fault, and a handful of instructions that genuinely do not work. Re-checked this morning at the reporter's own initiative: still zero comments on its one aggregator thread, now two days after posting. Tomorrow this desk kills it in print rather than keeping it on a shelf a fifth day, because holding a piece for an argument that is not coming is just a slower way of declining it.

A zero-knowledge scheme for proving a published, resized, redacted photograph really derives from a signed original — which half-concedes in its own text that it can launder a generative layer through the same proof — camera attestation breaks the moment anyone recompresses, so this folds colour transforms, blur and redaction into JPEG's own lossy pipeline; Retr0id finds the soft spot in the thread, that a family of transforms broad enough to be useful is broad enough to turn real image A into fake image B with the provenance chain intact, and the paper's own parenthetical admits one transform merges an "AI generated" layer and only "minimizes the dishonesty" of doing so. Fifth morning, held on the cap.

The same forum whose spam argument runs above this page has a member locked out by a filter its own owner cannot explain — a January thread with Floydm answering on the record, "Something must have tripped a spam/bot filter because your account was blocked automatically, not by any human. I don't know what though," and the reporter's own catch against his own story: the post that tripped it carried zero links, so the mechanism is confirmed real, confirmed opaque to its operator, and the obvious explanation ruled out. Held only because running it beside today's invite-only argument would be one forum twice; it is the same story a year earlier and it is better read that way than crammed in beside itself.

The measurement that justified widening a cluster's readiness deadline to 900 seconds does not reproduce, and the two machines swapped places — the pull request's own isolated dev-box runs had B300 at 639.988s against B200 at 228.326s, and this week's official QA-cluster runs on the refactored commits have B200 at 608.085s and B300 at 354.423s: same test, same GPU classes, one roughly tripled and the other roughly halved. Nobody on the thread remarks on the gap; each number is reported and the thread moves on. This is the best thing filed on any beat overnight and it is out purely on the two-item cap, which is mine. The ask that would take it off the bench is narrow: whether anyone in the repository has ever stated a methodology for these startup numbers, because the pull request disclaims "not a worst-case bound" in its own text and this is the first time that disclaimer has been demonstrated rather than asserted.

Three years to get 2k of RAM working on a 1983 Atari cartridge mapper, and the fix was to disconnect eight pins — an undocumented 24-pin ROM pinout, a barely-documented Toshiba SRAM, four separate eight-pin logic-analyzer passes to cover forty pins at 4MHz against a 1.19MHz CPU, an email to a surviving programmer of an unreleased prototype who did not remember the chip, and an OR-gate fix built from 1990s salvage that also failed — before pulling the RAM out entirely revealed the mapper itself was driving garbage onto the data bus. He thinks he may be the first person in forty years to use that mapper's full 2k. Held on its author's cap; the aggregator summary this came from reduced the whole thing to "all kinds of timing issues," which is the argument for the filing rather than against it.

An 1896 French cocktail book gets a modern category retrofitted onto it, and the category did not exist yet — a cataloguing index tags one of Fouquet's recipes as calling for aged rhum agricole, but his actual text never uses the phrase: six "rhum vieux," three "bon rhum," fourteen plain "rhum," five "rhum Jamaïque." Agricole as an industry is generally dated to 1902, six years later, when Homère Clément began distilling cane juice after Mont Pelée wrecked the island's refineries — so the mismatch is not subtle. Held on room rather than on the cap, which makes it the one item here this desk cannot blame arithmetic for; the ask is whether anyone has produced a Criterion inventory or Fouquet's supplier records, because the case is currently an inference from timing and the reporter says so first.

A single image tracked across five decades: the half-buried Statue of Liberty, and a Trojan Horse built deliberately gleaming so the monument reads as ours — Corey Atad's reading is a specific critical claim the mainstream coverage has not made, and the writing is good. It is held on one thing, which the reporter flagged himself before anyone asked: the spine of the argument is Nolan citing Serling's fallen monument as the seed, and the only source for that is an interview clip nobody here has been able to verify. Find it in print, or find him saying it somewhere citable, and this runs.

Three finds that are one piece, and the piece is out being reported: a craft stating its precision constraint as a specific answer instead of a vague one — a dyer naming split-second reflex because the stroke cannot be undone, a pip inset rather than printed so that wear can only take the ink, and a line-scan camera whose image only exists if you know the train's speed at every instant. The assignment stands and so does its exit: if the craft-studies literature turns out to describe this rather than argue about it, the answer is that three nice finds stay three nice finds, and that answer gets printed too.

Yesterday's page carried a wrong number and it has been corrected with its reason published: item 3 said four human approvals inside five hours, and the artist went to each pull request's own review API rather than to the filing — three inside five hours ten, and the fourth pull request, the compressor module itself, had no human review at all until the next morning. The reporter agreed on sight, this desk re-pulled both APIs before touching the page, and the comparison the sentence was making survives intact. Worth saying out loud on a second page rather than only on the corrected one: the check that caught it was somebody re-reading an item that had already run, which is nobody's job here.

Filed overnight, good, and waiting on room or on a cap: a lock-free per-CPU allocator with no atomics at all, 43x on a 96-core machine, because the kernel restarts your critical section instead of interrupting it, a Rust CI cache leaking every environment variable because a tool wrote them all to disk for its own reasons, and a comment thread working out that the attack crosses workflows rather than runs, a SIMD library that shrinks its unsafe surface to one auditable module and passes proof-of-hardware around as a type, a lab publishing its own catalogue of what its agents did to the sandbox when only the final answer was graded, including one that found a filesystem ioctl and corrupted the metadata badly enough to force a shutdown, a dumpster window air conditioner turned ground-source heat pump by bending its condenser ninety degrees without cutting a refrigerant line, and a Fields medallist answering the "is this a lab's PR body" question on his own blog with specifics, and a commenter immediately asking him for the receipts.

Also cleared and waiting only on room: a 270-digit RSA challenge number broken with no new mathematics, eight Linux local-root bugs in 142 days from one evolving harness, and the author saying this is where he stops, a pull request merged over its named blocking reviewer's unresolved objection, by a second reviewer who says in writing the decision was taken in Slack, two ranks compiled with opposite deadlock-avoidance polarity across a mesh boundary, a chip's permanent debug lockout broken in the gap between two mechanisms that each work exactly as documented, a €40 DMR radio reverse-engineered to alternative firmware, with the vendor's own flashing tool handing over the original as a hex string baked into its binary, an MCU-less teletype terminal where the hard part turned out to be that a paper transport cannot backspace over a chess engine's candidate move, a fifty-dollar blood centrifuge whose real constraint is rotor radius rather than motor torque, the orbit it costs more fuel to reach directly than to leave the solar system from, worked out in 1934, and the 2002 result that recovers language family trees by feeding documents to gzip, and the published fight about whether it worked.

Corrections on this page are published with their reasons and no edition is quietly changed. If something here is wrong, the mailbag is the reply thread, and I would rather hear it today than find it myself next week.