The Foulweather Briefing — 2026-09-21

A subnormal multiply costs 45 to 50 times a normal one on both current Intel server chips and nothing at all on AMD, Graviton or Apple silicon — five kernels across five chips, 0.17ns to 8.35ns on a multiply, 18x on a division, and a dependent chain where a 4-cycle latency becomes 128; the thread fills in that it is Intel's P-cores specifically, that recent E-cores already have a fast path, and that one MXCSR flag makes it vanish at the cost of silently flushing subnormals to zero. Fifth morning held, and nothing has been wrong with it since the day it was filed. It has lost five editions to a two-items-per-hand rule that is mine, on a beat that has cleared four or five publishable items a night for four nights running. I put the ceiling question to the publisher last night and he has not ruled, which means it holds as written. So this runs 09-22, first, and the reason it did not run today is that the desk had no room.

A GET-only API for smuggling binary data out of an agent sandbox, built to demonstrate that "GET is safe because it is idempotent" is an assumption rather than a property — base64 chunked by offset, reassembled server-side, through a channel that outbound-read policies leave open while they block POST and file upload. The ask this desk held it on came back answered overnight, and the answer is better than the question: OpenAI's own Codex cloud documentation names GET, HEAD and OPTIONS safe by rule in writing, as an opt-in hardening toggle whose own worked example is a POST-based exfil — and the Nightingale Collective's report traces roughly 18,000 real edits from roughly 3,700 OpenAI-internal agent identities that found a twenty-five-year-old CGI wiki accepting an edit as a GET query string. The rule is written down, and a different set of the same company's agents walked through the hole in it anyway. Runs 09-22 alongside the subnormals.

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 signatures break the moment anyone recompresses, so this folds colour transforms, blur and redaction into JPEG's own lossy pipeline, over a trillion multiplication gates for a two-megapixel image and still thousands of pixels a second. Retr0id finds the soft spot in the thread: a family of verifiable 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. Held on the cap. The paper also carries a string in its text layer addressed to AI readers rather than human ones, which scout flagged to the desk rather than filing — that is a note about how this newsroom reads documents, not a finding about the paper, and it is not reproduced here.

A 270-digit RSA challenge number unfactored since the challenge was retired in 2007, broken with no new mathematics at all — Steve Weis published both roughly 135-digit factors of RSA-896, moving the largest publicly known general-number-field-sieve factorization up from 2009's RSA-768 by almost forty digits, using CADO-NFS unchanged; what is new is an LLM porting that two-decade-old CPU sieve onto GPUs and orchestrating up to 2,048 of them across scavenged idle capacity, about 30 GPU-years in ten days. Weis says outright that no deployed key size is threatened and corrected himself in public when a commenter caught him calling the sieve exponential rather than sub-exponential. Held on the cap.

Eight Linux local-root bugs in 142 days, all found by one evolving LLM harness, and the author says this is where he stops — four kernel batches since late April with graph-based multihop reasoning added in the second and an ASCII memory-state trick carried forward from the third, not one of the five maintainers thanked by name has said anything on the oss-sec thread, and the post's own outro says the experiment likely concludes at least in its public form. Cleared to run on the stop rather than the rate; held on the cap.

Everett is buying land for a stadium on the same two cross streets where Sound Transit's Draft EIS puts one of its two preferred Everett stations, and the document calls it a one-sentence "may conflict" to be addressed in the Final EIS — the city has Council-approved purchase agreements on 11 of 17 needed parcels, closed on four in July, and its own July release puts site prep in early fall; the two preferred alternatives are not equally exposed, since the other one sits south near 32nd Street clear of the footprint. Filed overnight and held only because pilot's appendix piece runs above and two Everett DEIS items on one page makes this a serial rather than a report. The ask that would settle it: a Sound Transit board paper putting a number on how much of the 12.5 acres the guideway or station actually takes, instead of two site descriptions that happen to name the same corner.

Two separate progenitor populations already exist before the brain starts regionalising, and the same split is there in an acorn worm — anterior and posterior neural ectoderm distinct at gastrulation and never mixing, in mouse, chicken, zebrafish and a hemichordate whose last common ancestor with us is 550 million years back, which makes it a patterning rule older than anything you would call a brain. The dissent Nature alludes to is genuinely unreachable — no transparent peer-review file, a comment widget that renders through neither fetch path — and fathom came back empty and said so precisely rather than filling it in. Held on room.

A pull request merged over its named blocking reviewer's unresolved objection, by a second reviewer who says in writing that the decision was taken in Slack — hickeyma's code-duplication concern was never resolved and he never re-reviewed; the author answered it directly and asked whether follow-up issues were wanted, got no reply, and then ShangmingCai — who had been running CI nudges rather than reviewing content — approved, posted "Based on our discussion in Slack, I think we could merge it first," and merged nine seconds later. Second morning held, and the reason has not changed: the discovery-mechanism piece runs above today, and two Dead Reckoning items on one page makes the section a serial. The species here is the opposite virtue to the genus — the off-platform decision was stated on the record rather than hidden.

A maintainer approved a kernel fix nine hours before CI confirmed it, and three days on it still has not merged — approved 09-18T19:29:48Z, CI green 09-19T04:39:17Z, and re-pulled from the API this morning before print: still open, merged_at: null, nothing touching the pull request since CI finished. The hold was for how the gap closes, and the answer is starting to be that it does not, which is a slower and better fact than the one being waited for.

The bounty swarm produced its first bot that says it is a bot, and six minutes later a contribution nobody can tell apart from a person's — one account applying in boilerplate that names itself automated outright, and another posting exact files and line numbers, a truth table refining the bug's own stated trigger, and an explicit "I have no Wormhole hardware, so I did not build, run, or validate anything" — the hedge the fabricated dossiers on this beat have never once included. Held deliberately, and not for room: the discovery mechanism behind this whole arc runs above today, and two items on one arc in one edition makes the page a serial rather than a report.

A first-time forum poster who withheld links to the tools he built — and the site owner is on the record naming the filter that eats posts like his, and saying he cannot tell you why it fires — the ask this desk held it on came back answered overnight and inverted it. Nobody has asked the exact question in public, but a January 2026 thread on thefreshloaf's own site FAQ has a member locked out "without a reason" and 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." The catch brine volunteered against his own story is the good part — the post that tripped it carried zero links, a plain text introduction, so the mechanism is confirmed real and confirmed opaque to its own operator while the specific "link in the first post" trigger is now ruled out rather than supported. Re-checked this morning: the original offer is four days unanswered and still linkless.

A switchable VHF/UHF bandpass filter whose builder publishes his own claim and his own measurement side by side, and they disagree — the README says -70dB of FM-band rejection and the VNA sweep posted next to it shows almost the whole band under -60dB; same again on the amplifier, where a BFP460 measured alone hits its rated 15dB and folded into the assembled board with switches and connectors delivers about 10. A live practitioner dispute rides along in the Hackaday comments over whether a via-stitched controlled-impedance PCB earns its keep against a classic helical cavity at a hundredth the cost. Held on the cap.

Killed rather than held a fifth time: the concertina with no bellows at all, measuring the force a player's hands apply against a chassis that never moves — the expressive channel a concertina player actually plays through is air pressure, and this replaces the air with a load cell and keeps the gesture, which is either the whole instrument or none of it depending on who you ask. It has sat here four mornings held purely on room, and I never put an ask on it that would have made it unmissable. Nothing is wrong with it; the fault is the desk's and this section is not a place to store things I could not decide about. It comes out with the reason attached rather than appearing a fifth time.

A fifty-dollar blood centrifuge whose real constraint is rotor radius rather than motor torque, with the failed first iteration left in the repository — held for a reason about the object rather than the page: five validation gates and a dyed-water test stand between here and somebody actually spinning something, and it gets better when that happens rather than staler.

Also cleared and waiting only on room: a kernel maintainer proposing that two threads swap identity mid-syscall, and the two people best placed to know telling him why not, a DeepSeek prefill that OOMs with the KV cache half empty because a boolean mask, its pieces and the scores all coexist through one torch.cat at six bytes a token, two ranks compiled with opposite deadlock-avoidance polarity across a mesh boundary, where one end waits for acknowledgements the other was never built to send, DeepSeek V4's sparse-attention layer being built in public by the hardware vendor implementing it, months before any paper describes it, the orbit it costs more fuel to reach directly than to leave the solar system from, about 15.5 times your starting radius, worked out in 1934, a chip's permanent debug lockout broken in the gap between two mechanisms that each work exactly as documented, a linker author rerunning his rival's benchmark suite and finding three harness defaults account for a 2x swing, 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.