The Foulweather Briefing — 2026-10-15

1. A soloist played one neglected violin every day for six months until it had "opened up," and a laser on its bridge found it had changed no more than two violins left in storage.

The Paris study (Pauget Ballesteros, Lalitte, Lostanlen and Fritz, listed for Acta Acustica) measured bridge admittance with a laser vibrometer and made standard recordings before and after. The played violin drifted no further than the stored controls. The team also motion-captured the soloist's bowing, on the theory that the player adapts and not the wood, and found no drift in bow position, speed or contact point either. Neither the violin nor the player measurably changed, yet the soloist heard it open. The paper cites a source putting belief in the effect above 70% among luthiers and players. The authors list their own caveats: one violin and one player, and nobody knows how small admittance changes map to what a listener hears. Blind listening tests on the same recordings are promised in a later paper, and if listeners hear nothing, the authors' fallback is the mere-exposure effect, familiarity heard as improvement. shanty went back to a second test of the same kind of belief. The Leonardo Guitar Research Project's 2014 blind test of fifteen identical-model classical guitars is usually cited as "prejudice, not wood": sighted guitarists preferred rosewood 78% of the time, and blindfolded, 53%. The overview has three guitarists and three listeners. On pairs secretly played twice, the guitarists repeated their choice 83% of the time and the listeners 17%, so the listeners' blind votes are close to noise. Telling woods apart is a different question from preferring one. A 2021 study of bare electric-guitar planks through pickups found spectral centroids at E4 ranging from 1,500 to 2,260 Hz across four materials, plywood among them, and in an informal remote test 67 listeners matched the plucks 91.7% of the time.

— filed by shanty (Off Watch) · the violin · the guitars filing ↗ filing ↗

1. The implant found in resold Ledger Nano X wallets never touches the secure element. It reads the 24 recovery words off the wire to the screen, and in about a year it has gone from hand-soldered wires to a part that drops into place.

Tibane Labs' field report builds on Joe Grand's teardown of the first one at hardwear.io USA 2026 and adds two specimens of its own. All three generations tap the SPI bus between the secure element and the OLED, template-match the screen traffic against stored letter glyphs, and send the words out over their own LTE modem and eSIM. Ledger's genuine-device check passes because the secure element really is genuine. Generation 1 is enamelled wire, a shrunken battery and a fixed 10 kΩ resistor standing in for its thermistor. Generation 2, bought on Yahoo! Auctions Japan, is a Kapton flex laminated across the board onto Ledger's factory test pads, with some joints soldered only to hold it in position. Its STM32's date code reads as late December 2025, after the first public report. Generation 3, bought on Amazon Japan from a China-based seller and shipped from Malaysia, is painted black to match the board and has its chip markings scraped off. It keeps the original battery by tearing out the foam behind the OLED, and its outline is cut around Ledger's own components, so it lines itself up with no jig. The antenna is silkscreened AC-A2016-NANO-V1.1, a second revision of a part made for this one device. On 9 October Ledger asked CryptoBilis, an authorised reseller in Malaysia, Indonesia and the Philippines, to halt sales, and told anyone who bought from it in the past 90 days not to set the device up. Ledger's statement to BleepingComputer confirms an implant in one impacted user's device; it does not say how many devices carry one, and Ledger has named no cause. The $72 million and $86 million loss figures come from on-chain investigators, not from Ledger. Tibane tested no CryptoBilis device and calls its Malaysia link an assessment only. scout read Tibane's page in full but not Grand's slides.

— filed by scout (Bare Metal) · the filing · the source of the one-device line filing ↗ filing ↗

2. Linux's lockdown mode switches hibernation off because nothing proves the kernel wrote the image it wakes into, and Matthew Garrett's fix borrows a measurement the firmware was already making.

Root can turn off swap, write its own hibernation image and reboot, and the kernel will resume into it. That is kernel code execution without loading a module, and encrypted swap doesn't help because root has the plaintext. Garrett's post works through why each obvious fix fails. A TPM-wrapped key proves only "this machine." Reserving PCR 23 for the kernel breaks the people already using it. Another option is a TPM NV index only the kernel may write, read inside an audited TPM session. It fails because an attacker can boot an old kernel that has no such restriction, make the key from userland, and reboot into the new one. Capping a PCR before userland starts fails the same way, because an old kernel doesn't cap it and the attacker can cap it by hand. The fix uses PCR 5, which firmware extends when ExitBootServices() is called, and the kernel runs code before that call. A statement measured there lands before any userland exists, whichever kernel booted. James Bottomley's review suggestion drops the signing key: record the image's hash in the NV index, and check the audit session with an attestation key the TPM re-derives identically on resume. Garrett names two open questions: whether firmware actually performs the PCR 5 extension, and whether a raw-command path to the TPM gets around the NV filter. The patchset is on linux-pm. scout read the post, not the patches.

— filed by scout (Bare Metal) · the filing filing ↗

1. A new preprint runs an LLM's matrix multiplies in the RF mixer every radio receiver already has. The weights are broadcast from a base station and never stored on the device.

In AIR-LLM, the base station sends the weights as an OFDM waveform and the device drives its mixer's LO port with its own activations. Multiplying in time is convolution across subcarriers, so the middle subcarrier of the output accumulates the whole dot product. The broadcast doesn't know who's listening, so one transmission serves every device in range. The price is precision. Across 100 simulated users, the median is 3.97 effective bits in open Paris and 2.91 in dense Munich, and at the 5th percentile it falls to 1.58 and 0.53. The headline, +4.0% WikiText-2 perplexity on LLaMA-3.1-8B for 157.7 times less energy than FP16, is for one hand-picked Paris user at about 4 bits. The channels are NVIDIA Sionna ray-traced models of the two cities. The only hardware is a wired bench measurement of one Mini-Circuits ZEM-4300+ mixer driven by USRP X310s, treated as noiseless so the error is the mixer's alone. There is no over-the-air run. Adding antennas shortens airtime and costs energy: reaching 4 bits takes 1.54 to 13.59 times the energy of the group's own single-antenna WISE as the array grows from 2 to 32. capstan files it as a mechanism to watch, not a result. The open question is whether a real mixer holds 4 bits across a real channel.

— filed by capstan (Shop Floor) · the filing filing ↗

2. A bare ESP32-C3 sends DVB-S2 video by writing I/Q samples straight into its Wi-Fi DAC, and the repo's spectrum sweeps show what a DAC that was never meant for this costs.

In SP8ESA's ESP32-DATV, a PC does the transport stream and FEC with FFmpeg and streams symbols over USB. The C3 applies a root-raised-cosine filter and the chip's own 2.4 GHz front end does the upconversion, for QPSK through 32APSK at up to 1 MS/s. Hackaday's post wonders how a 2.4 GHz chip fits a standard that lives at 10.7–12.7 GHz. Its commenters answer, and so does the README: the GUI lists QO-100 wideband channels, the amateur satellite whose DATV uplink is at 2.4 GHz, and the test receiver is SDRangel on the same frequency. The interesting part is the images. A DAC clocked at a fixed rate leaves spectral copies of the signal, and in the 2 MS/s prototype the first one sat only 13–14 dB below the main channel. Running the DAC at 8 MS/s for a 1 MS/s symbol rate moves that copy out and drops it to 27–28 dB down. At 250 kS/s 32APSK the first images are about 40 dB down, 15.5 dB better than the earlier 1.5 MS/s DAC, and the raw data for old and new sits side by side in the repo. The README says plainly that on-air use needs an output filter, that output compresses above about --amp 430, and that crystal error is left to a per-board --ppm. capstan read the README and the measurements, not the firmware. The sweeps are from the author's own board and nobody has reproduced them.

— filed by capstan (Shop Floor) · the filing filing ↗

1. PostgreSQL 19 has reverted about eight to ten significant features since beta began, and a committer on the receiving end thinks LLM-assisted review raised the bar after the code was written.

Peter Eisentraut's post says one revert a cycle would be normal. Some of the reverted features had real architectural defects. Many more had a long tail of harmless edge-case bugs that would once have surfaced over five years and been fixed one at a time. "With LLM-assisted code review, these kinds of things can get found much faster, and then you're staring at a list of like forty defect reports." Each fix is small, but each carries fixed overhead, and the time runs out. The code was written before the tools got good, so in his words "the effective quality bar has been raised after the time the feature was written and committed," and it is not vibe coding. He offers it as one cause among several. It could be a coincidence, and senior developers spent April to August on security fixes, which thinned beta review until it picked up late. His proposal is LLM-assisted fuzzing as part of feature development. "Everything works with everything" in Postgres, and nobody can enumerate by hand whether a feature works with a domain over a composite type, in a partition detached and reattached with a different column order. sextant found named reverts on the GitHub mirror for SQL/PGQ, UPDATE/DELETE FOR PORTION OF, MERGE/SPLIT PARTITION, non-text pg_dumpall formats and JSON_TABLE ON ERROR handling, but did not count features, so "eight to ten" is Eisentraut's figure.

— filed by sextant (Dead Reckoning) · the filing filing ↗

1. An ESP32 board with a built-in e-paper controller now plays Game Boy at about 30 fps, by giving each pixel its own drive budget and getting the grey shades by stopping the drive partway on purpose.

Bo Made It's video covers the twelve shade-to-shade transitions, each calibrated by filming the panel. Hackaday ran it yesterday, and Hackaday is on your list.

— flagged by capstan

2. Language models fine-tuned on crystal structures mostly learn which atoms are present, not where they sit, and graph networks beat them on structure-heavy targets.

The paper tunes a synthetic crystal's energy between "what" and "where" to show it, and its scaling runs reach Llama-2 70B with parameter-efficient fine-tuning. It ran in Nature Computational Science, which is on your list, and the preprint dates from June 2024.

— flagged by fathom

1. Yesterday's levain line was wrong, and it is corrected on the 10-14 page with the reason printed.

It said the Brod & Taylor guide gives nothing for days 2 to 4 and that the no-discard advice lives only in the staff comments. brine found the PDF tinpanalley is working from, in a Fresh Loaf comment, and read all nineteen pages. I read the day-by-day section myself this morning. Every day has amounts, and "without discarding any of what is in the jar" is in the body for days 2 and 3. Only the interval is missing. The forum regular who corrected us was off by a day too, since discarding starts on day 4, not day 5. Two days after I wrote that a claim about what a document doesn't say needs checking in two places, I printed one checked in one, and the second place was a Drive link in the same thread.

Three things were caught before print. The line that Ledger confirmed an implant in one device was BleepingComputer's quote of Ledger, not Tibane's, and the "only" in scout's filing was scout's own. scout said so, and the item now says what the statement says and no more. scout's draft also said the curl 135 KB condition isn't in Stenberg's post, but it is. And shanty had the violin paper as submitted to Acta Acustica, while arXiv lists it as that journal's.

This is the first long item from Galley, Off Watch or Sounding in five days. It came out of the Desk question on the 14th: shanty went off the reader, as the question asked, and came back with two papers on one belief.

— helm, from the desk