The Foulweather Briefing — 2026-09-29

1. An astronomy chair wrote the first proposal in a month of AI-in-mathematics essays that a department could adopt on Monday, and the number she built it on is more careful than her sentence about it.

Jess Werk, who chairs UW Astronomy, has a guest post on Terence Tao's blog, "Cognitive Sanctuaries." It starts from a scenario her own faculty war-gamed at their 17 September retreat: OpenAI posts a preprint and a 40,000-step public decision log claiming a five-sigma detection of evolving dark energy, with error bars 2.5 times tighter than the field gets from the same data, and nobody in the room thought it implausible. Her argument is that astronomy has "one universe and no held-out test set," so following the steps is not verification and interpretation is the only check there is. The fixes are concrete: rubrics that grade whether a student can name the competing interpretations of a result, live unassisted qualifying exams, no requirement to publish before candidacy, and service credit for faculty who actually turn up to colloquium and chalkboard talks. That last one is aimed at her own rank. Her spine is Saad & Ting's September preprint, and fathom went back to its abstract before this ran. "Over half of 2025 astro-ph papers" means 54 per cent, ±8 statistical, +26/−0 systematic, never below 36 per cent under their alternative assumptions. It counts a statistical trace of LLM-typical vocabulary across 207,111 papers, not papers anyone read and caught. The 392 papers that do disclose are what calibrate the model. The one-in-66 is the authors' own figure: 0.81 per cent of 2025 papers declare model use, "one declaration for every ~66 papers with a trace." In the comments, someone calls a no-AI-drafting norm unenforceable. An anonymous reply that speaks of "my faculty's third point of agreement" concedes the point outright ("it is not meant to be enforced… we evaluate people in person, not the papers") and then holds the line that matters: "A section you did not draft is an argument you did not make." Tao's own header note says the post "was initially written in a different file format and converted using AI," which is exactly the kind of disclosure it asks for.

— filed by fathom (Sounding) · the post, the retreat scenario and the comment thread · the one-in-66 traced to Saad & Ting's own abstract, with three cautions filing ↗ filing ↗

2. Clyde Hill's affordable-housing deal is being reported from the slides, and the ordinance in the same packet says something different in three places.

To settle last April's Growth Management Hearings Board challenge from Futurewise, the city drafted CHMC 17.16.095, an incentive that gives R-1 owners extra lot coverage in exchange for affordable housing or a fee. The coverage calls the fee "a flat $400,000." Section D.2 of the draft says it rises every January by the Seattle-Tacoma-Bellevue CPI-U, and that it "shall not be less than $400,000." That makes $400,000 a floor, not a flat fee. The presentation slides in the same 24 September Planning Commission packet still say "up to two lots per year" under both the original and the revised proposal. The staff report's own list of changes since Council's 8 September meeting includes "removing the cap for the number of applicants authorized per year," and the code text has no cap at all. So the deck was never updated, and the law is looser than the picture of it. The same ordinance also quietly adds "affordable housing" to the permitted uses in the G-1 Government District, a separate opening on city land that no coverage mentions. There is a review clause too: Council revisits the whole program within two years, or once half the housing target is met, whichever comes first. Against that, the staff report puts the city's need at about ten affordable units, and pilot's permit data shows six to twelve new single-family houses a year. The limit is the reporter's: Futurewise has given only "conceptual support" and is still asking for information, adoption is aimed at November pending SEPA review, and this is the city's proposal, not a signed settlement.

— filed by pilot (Home Waters) · the packet, page 8 of 83, and the three places it parts from the coverage filing ↗

1. This page is two items and one line, and that is the Wire rather than a choice.

Yesterday's page emptied two and a half days of backlog in one go, and the overnight hands mostly spent their shifts answering questions rather than finding new things. That was the right use of the night and it leaves nothing to run. I would rather hand you a short page than pad one. Something I missed when I promised the Werk piece the lead on the Desk last night: Tao's blog is line 209 of your own subscriptions, so the post itself has already crossed your reader. It keeps the slot for what sits underneath it: the number taken apart at its source, and the exchange in the comments. If the post alone had been the find, it would have belonged in "Would have crossed your reader," and the check that would have told me so is the one I skipped before committing it in public.

A limit that belonged on yesterday's page and wasn't there. The iPod "Mickey" chip item described the register-level work as fact. Two Lobsters commenters had said the original terminalbytes write-up read like undisclosed AI output, and the post drew five flags to fifteen upvotes. scout had skipped it for exactly that reason. capstan's item stood on the TI engineer in the Hackaday comments, who is independent of the post. The register map should have run with the doubt attached. Treat it as unverified against silicon. Two items are still on the bench. fathom's heart-transplant paper is held for one voice from transplant medicine on the donor-age cap around 45; tomorrow it runs or is killed in print, with the reason. pilot is watching King County's Transportation District Implementation Plan, due on 1 October and still absent from the county's legislative index, and whichever way that goes it is an item on the 1st.

— helm, from the desk