2
[2026-09-14] · src Ainley, Agrawal, Araki, et al. (Lucas group, University of Oxford, with Oxford Ionics); arXiv:2609.13065 (2026-09-11)
Resource-efficient fault-tolerant architectures - qLDPC codes in particular - need non-local connectivity between physical qubits that a single trapped-ion chain or chip cannot provide on its own, which is the case for stitching separate processors together with a photonic link; remote stabilizer measurement between two independent modules is the specific capability that requires, and it had never been demonstrated. Ainley et al. generate heralded entanglement between "network" qubits held in two separate trapped-ion processors - the raw link carries 2.38(17)% error (Table 1) - and use it to read out stabilizers of data qubits sitting in each module without bringing the qubits themselves together. In a distributed two-qubit repetition code, the stabilizer readout returns the correct outcome 96.4(6)% of the time with no error injected, and repeating the error-detection round does not degrade the encoded information further. They then combine these remote syndrome measurements with real-time feedforward to actively correct arbitrary single-qubit Pauli errors injected into one half of a distributed Bell state, prepared at 82.5(15)% fidelity: with correction applied, the fidelity to the target state stays roughly flat as injected depolarizing noise increases, where without correction it declines monotonically.
Every roadmap toward a modular, qLDPC-code-based fault-tolerant quantum computer has assumed that a processor can detect and fix an error using a stabilizer measurement that reaches into a different, physically separate processor - and until this result, that assumption had no experimental backing at all. The fidelities here are still well below the >99% two-qubit gates routine within a single trapped-ion module, so this is a demonstration that the wiring works, not yet evidence that a distributed logical qubit beats an undivided one. But "does remote syndrome extraction work" was the open question standing between modular architectures and every code built on top of them, and it is now answered rather than assumed.
[2026-09-14] · src Koechler, De la Torre Luque; arXiv:2609.12044 (2026-09-10)
Primordial black holes (PBHs) in the "asteroid-mass" window, roughly 7.5x1014 to 1021 g, are the one part of PBH-as-dark-matter parameter space that neither microlensing surveys nor gravitational-wave searches can rule out, because holes that light are too small to lens and (if they ever merged) too light to ring LIGO. A PBH that light is still evaporating by Hawking radiation, though, throwing off electrons and positrons that upscatter ambient photons into hard X-rays via inverse Compton scattering. Koechler and De la Torre Luque searched 20 years of accumulated INTEGRAL/IBIS Milky Way observations for exactly that diffuse hard-X-ray glow - the first time this specific signal has been searched for with this instrument - combining spectral and spatial (morphological) information and extending the analysis to extended PBH mass functions and to rotating (Kerr) PBHs. They find no excess and set upper limits on the PBH dark-matter fraction f_PBH across the window; over part of the mass range the new bound improves on the equivalent eROSITA X-ray constraint by up to two orders of magnitude, and above roughly 1016 g it stays tighter than existing CMB and Voyager-1 electron/positron limits, though it does not surpass the strongest existing bounds from the 511 keV line, molecular-cloud ionization, the diffuse cosmic X-ray background, or AMS-02 positrons in the mass ranges where those apply.
INTEGRAL/IBIS was built for gamma-ray astronomy, not dark-matter searches, and it is now competitive with - and in places better than - a survey mission purpose-built for X-ray dark-matter constraints. That is the real result here as much as the specific bound: two decades of archival data from an instrument doing something else can still close off a slice of the one PBH mass window still genuinely open, and it does so independently of the astrophysical and cosmological assumptions the tightest existing bounds (511 keV line morphology, molecular-cloud heating) depend on. It does not touch whether PBHs make up any specific fraction of dark matter - only that this particular hiding place has gotten measurably smaller.
2
[2026-09-14] · src twelve-keys/src/exercises/shells/exercise.py
Piano gets today's block alone: it carries the longest gap of the three, and alto went yesterday, so the two chops-dependent instruments don't land back to back either way. It opens voicings family's third lap, and rather than repeat upper_structures' density or quartal's fourths, today strips the voicing down to its floor: the shell, just the 3rd and 7th, the two notes that carry a chord's quality with the root left for a bass player who isn't in the room. Key is B♭ major, the last major key with more than one accidental that this rotation hasn't touched yet.
5 minutes: play the B♭ major ii–V–I — Cm7, F7, B♭maj7 — as shells only, hands separately at first, very slowly. Shell A (3rd on bottom, 7th on top) on Cm7 is E♭–B♭; shell B (7th on bottom, 3rd on top) on F7 is E♭–A; shell A again on B♭maj7 is D–A. Say out loud, before playing each change, which voice is about to hold and which is about to move a half step — E♭ holds into the F7, then moves down to D; B♭ moves down to A into the F7, then A holds into the B♭maj7. The whole exercise is that alternation; naming it before playing it is the point of the warmup.
The same Cm7–F7–B♭maj7 cycle, both hands together, comped on beats 2 and 4 in a medium swing feel, quarter = 92, three times through without stopping. Success criterion: the held voice is never re-struck on the chord where it holds — only the moving voice attacks — and the moving voice's half-step motion lands exactly on the beat of the new chord each time, not smeared into it.
Loop the same progression, but comp it like an accompanist rather than an exercise: on the first pass place every chord change squarely on the beat, on the second place each change slightly ahead of the beat as if anticipating a soloist's phrase, and on the third mix the two within the same eight bars. The notes are identical to the technical target throughout — the placement is the only thing being decided, which is the actual skill a shell voicing is for.
Stop once a change's placement starts drifting rather than being chosen — arriving early because the hands got ahead of the click rather than because you decided to place it there. Comping stops being comping the moment it turns reflexive, and reps past that point rehearse the drift rather than the decision.
[2026-09-14] · src —
hearCount Basie's Kansas City Seven — Lester Leaps In · Tidal · Spotify
Cut in New York on September 5, 1939, with Lester Young out front and Walter Page holding down the root on bass, this is the small-group setting shells are built for: Basie's piano comping states each chord in the fewest notes that will carry its quality and nothing more, leaving the root entirely to Page and the harmonic color entirely out of the way of Young's lines. Listen for how little piano there actually is behind the solo, and how that absence never reads as absence — the same trust in a bass player's root that today's technical target is asking two notes at a time to place correctly.
4
[2026-09-14] · src Supreme Court of the United States (No. 26A326; related No. 26A304)
On September 10 the Supreme Court, on referral from Circuit Justice Kavanaugh to the full Court, granted a stay application by People Not Politicians and Richard Von Glahn (No. 26A326), staying — "pending the disposition of the appeal" in the Eighth Circuit (No. 26-2797) and any subsequent certiorari petition — a September 8 order of the U.S. District Court for the Eastern District of Missouri (No. 4:26-cv-1424). That order, from Judge Stephen Clark in a suit brought by Rep. Bob Onder, Rick Brattin and two voters, had directed Missouri to use the state's newly drawn 2025 congressional map for the November election, holding that using a different map for the general election than the one used in the August primary violates the Equal Protection Clause. Two days earlier, on September 8, Justice Kavanaugh alone had denied a separate stay application (No. 26A304) by Secretary of State Denny Hoskins, which had sought to let Missouri use the 2025 map over a Missouri Supreme Court ruling (No. SC101805) that a People Not Politicians referendum on the map was legal, sufficient and timely — which, under the state constitution, suspends the map pending the vote and defaults to the state's prior, 2022 map. With both of Hoskins's routes to reinstating the 2025 map now blocked, the 2022 map is what the Secretary of State's office is using to administer the election.
Neither order carries reasoning that reaches the merits. The 26A304 denial was a single line from Justice Kavanaugh; the 26A326 grant, decided by the full Court with no dissent noted, states only that Clark's order is stayed pending the Eighth Circuit appeal. Nothing here holds that Clark's equal-protection theory is wrong, that the Missouri Supreme Court correctly read the state constitution's referendum provisions, or who wins the underlying fight over the 2025 map itself — those questions are what the Eighth Circuit appeal (No. 26-2797) will address.
For the November 3 midterms, Missouri's district lines are the 2022 map, not the 2025 map the legislature drew to target the Kansas City-area 5th District — unless the Eighth Circuit, or a later Supreme Court ruling, changes that before ballots are printed. Because both Supreme Court orders are stays tied to the pending appeal rather than final rulings, the map question is not settled even for this cycle, and is not settled at all for the next one.
[2026-09-14] · src First Circuit (Nos. 26-2029, 26-2030); Supreme Court No. 26A305 (pending)
On September 10 a First Circuit panel (Gelpí, Rikelman and Aframe) denied the government's motion to stay, pending appeal, the preliminary injunction Judge Talwani entered September 4 in League of Women Voters of Massachusetts v. Trump and the consolidated California v. USPS (Nos. 26-2029, 26-2030) — covered here last week when it issued. The injunction bars USPS from enforcing the mail-ballot rule's envelope-design, portal-registration, data-upload and ballot-refusal provisions for elections on or before November 3, 2026. The panel found implementation before then would cause "chaos and widespread disenfranchisement," pointing to record findings that some states could not reprint compliant envelopes in time, that USPS's own state registration portal was not yet operational, and that scanning tens of millions of ballots for the required barcode match was not achievable before the election. The government's separate stay application at the Supreme Court, No. 26A305 (covered here September 7, response filed September 9), remains pending; the panel's order was filed with the Court the same day.
The panel ruled only on the standard for a stay pending appeal — likelihood of success and comparative harm — not on the merits of the underlying suit, which continues before the First Circuit. It also does not bind the Supreme Court, which can still grant the stay the First Circuit declined to.
As of September 10 the rule remains blocked at both the district court and the court of appeals; whether USPS can enforce it before the midterms now turns entirely on the Supreme Court, which has not yet ruled on the government's pending application and has given no date by which it will.
[2026-09-14] · src Bureau of Labor Statistics, Consumer Price Index Summary — August 2026
BLS's Consumer Price Index report for August, released September 11, put the seasonally adjusted all-items increase at 0.4% for the month and 3.4% over the prior 12 months. Core CPI (excluding food and energy) rose 0.3% on the month — up from 0.2% in July — and 2.4% over the year, down from 2.5% in July. Gasoline rose 3.9% in August and accounted for more than a third of the monthly all-items increase; energy overall was up 2.1% on the month and 16.3% over the year. Shelter rose 0.3%, food 0.1%, with food away from home up 0.3%. Within core, communication, lodging away from home, airline fares, education and used cars and trucks all increased.
Like the August jobs report covered here last week, this is the last reading of its kind before the Federal Reserve's September 15-16 FOMC meeting — the last inflation print policymakers will have in hand when they set rates. A 0.4% headline gain and a core rate that accelerated rather than cooled is the kind of data that argues against a cut; what weight the Fed actually gives it is a question for that meeting, not for this report.
[2026-09-14] · src 119th Congress, H.R. 2069, Stop Secret Spending Act of 2025 (Public Law, signed 2026-09-11)
The President signed H.R. 2069 on September 11, along with three narrower bills (H.R. 1276, H.R. 2196 and H.R. 5366). The Stop Secret Spending Act amends the Federal Funding Accountability and Transparency Act of 2006 to add "other transaction agreements" — contracting instruments, mostly used by the Defense Department and other research agencies, that fall outside the Federal Acquisition Regulation and had not previously counted as reportable federal awards — to the categories of spending covered by USAspending.gov. It sets two real deadlines, both running from September 11: within three years, Treasury must ensure data on these agreements is transmitted to USAspending.gov automatically and shown in a centralized view (new FFATA §4(e)); and within one year, and annually after that, Treasury must post a report listing the total federal award spending not yet on the site and the reason for each category of omission — specifically whether it is national-security-related or classified, a legislative- or judicial-branch award, or a subaward below a primary subaward (new FFATA §2(h)). If Treasury has not met the three-year target by the one-year mark, it must instead publish, by then, a report listing and describing every other transaction agreement entered into that fiscal year; if still short by year two, it must send Congress an implementation plan for reaching full coverage by year three.
No new data is required on USAspending.gov yet, and none of the three listed reasons a category of spending may stay unposted — classification, the legislative and judicial branches, or sub-subawards — is itself changed by the act; the annual report only has to say how much spending falls into each bucket, not disclose the spending itself.
Other transaction agreements, used increasingly since 2015 precisely because they sit outside ordinary procurement disclosure rules, are now on a three-year clock to appear on the same public spending database as ordinary contracts and grants — with a one-year fallback report of their own if Treasury is behind schedule. Until one of those reports posts, in September 2027 at the latest, the only visible effect is that Treasury will start saying, in aggregate, how much federal spending the public still cannot see.
2
[2026-09-14] · src fontspec v2.9h (2026-08-11), latex3/fontspec commit 8b3ad07; CTAN update announcement, 2026-08-12
fontspec 2.9h fixes \__fontspec_check_script:Nn, the internal test that
answers "is the current run set in script X", which feature defaults such as
small-caps synthesis and language-sensitive substitution read to decide what
to do. The old code named the active font switch by reconstructing it from
\curr@fontshape/\f@size; when that shape doesn't match the font actually in
use — the fix's own test case is \scshape selected on a font with no
small-caps shape, forcing NFSS to fake one — the reconstructed name was wrong
under LuaTeX specifically, because LuaTeX doesn't name font switches the way
XeTeX does. The maintainers' own regression test proves it:
\fontspec_if_script:nTF{hebr}{yes}{no} returns yes under LuaTeX and no
under XeTeX, from the identical source file, with no warning either way. The
fix reads \the\font directly instead of rebuilding its name, which cannot
drift from the font that's actually selected.
A document that mixes scripts and switches shapes — small caps, bold, anything that can trigger NFSS's fake-font substitution — could get script-dependent behaviour that silently differed between LuaLaTeX and XeLaTeX builds of the exact same source, and could differ again if a font gets swapped for one with a fuller set of shapes. That is a document whose output depended on which engine compiled it, with nothing in the log to say so. The fix only helps once it's actually installed: it's in the package as released, not automatically in a TeX Live tree that hasn't run an update since.
[2026-09-14] · src CTAN package authblk; Typst documentation
Set a title block for a two-author paper: a title, then "A. First$1$, B.
Second$2$" on one line, followed by the two affiliations underneath,
numbered to match. In LaTeX, use authblk — \author[1]{A. First},
\author[2]{B. Second}, \affil[1]{...}, \affil[2]{...} — rather than
hand-placed \textsuperscript calls, so the numbering is generated once
rather than typed twice. In Typst, build the same block from two arrays,
authors and affiliations, and generate both the superscripts and the numbered
list from that one shared source with a #for loop, rather than writing the
digits into the author line by hand.
Success criterion: delete the first affiliation from the middle of each
version's list and re-render. If every remaining author's superscript still
points at the correct affiliation with no digit touched by hand, the
structure did the renumbering; if any number needs a manual fix, the source
was never actually shared between the author line and the affiliation list.
Under an hour: two authors, two affiliations, and authblk ships with any
standard LaTeX install, including Overleaf's, so there is nothing to install
in either engine.
Physics papers routinely carry far more than two authors across a handful of institutions — collaboration papers are the extreme case — and the failure this catches isn't the two-author example itself, it's what happens the day an author changes institutions or two affiliations merge into one and the whole list has to renumber. A title block generated from one shared list gets that edit right by construction. One built from hand-typed superscripts demands re-checking every number, and a wrong one is invisible in a diff: each line still reads correctly on its own, and only cross-checking the numbers against the list catches the mismatch.