4
[2026-08-09] · src arXiv:2407.10913 (hep-lat/hep-ph); published as Nature (2026), doi:10.1038/s41586-026-10449-z
A hybrid calculation of the leading-order hadronic vacuum polarization (HVP) contribution to the muon's anomalous magnetic moment — lattice QCD for the short- and long-distance windows, e+e- → hadrons cross-section data for the window in between — reaches 0.48% precision and predicts a_mu within about 0.5 sigma of Fermilab's final measurement. That's a near-total collapse of the ~4 sigma gap that has stood since 2021 as the best hint of physics beyond the Standard Model. It doesn't erase the disagreement so much as relocate it: the paper's lattice-based intermediate-window value differs from the KLOE-based e+e- value by 6.2 sigma and from BaBar's by 3.5 sigma, while sitting close to the newer CMD-3 measurement.
The 25-year-old muon g-2 "anomaly" was never a single number — it was a comparison between one experiment and a Standard Model prediction built mostly from decades-old e+e- → hadrons cross-section data. This result says the muon's magnetic moment matches the Standard Model after all, but only because at least one of the historical collider datasets that prediction was built on (KLOE, and to a lesser extent BaBar) now looks wrong at high significance, in a way nobody has explained yet. New physics didn't rule itself out here so much as get displaced onto a mundane-sounding but unresolved systematic-error hunt in 20-year-old hadron cross-section measurements — a correction of the field's own inputs, not a discovery.
[2026-08-09] · src arXiv:2604.05746 (SNO+ Collaboration)
Using 685 live-days (May 2022-July 2025) of scintillator data at SNOLAB (Sudbury, Ontario, 2 km rock overburden), SNO+ reports the first direct geoneutrino flux measurement outside a Northern-Hemisphere continental craton: 49(+13/-12) TNU (terrestrial neutrino units, events per 1032 target protons per year), consistent with reference geological models of U/Th decay in the crust and mantle. Every prior geoneutrino measurement — KamLAND (Japan), Borexino (Italy) — sat on the same broad crustal province.
Geoneutrino counts are the only direct, model-independent probe of the radiogenic heat (from U and Th decay chains) that drives mantle convection and plate tectonics, but disentangling the crustal contribution from the mantle contribution has always leaned on a single local crust model, because every prior detector sat in roughly the same geological setting. Sudbury is a different craton with an independently constrained local geology, so this measurement is a genuine second data point for cross-checking the crust/mantle split rather than a replication of the same one. The central value matches expectations, so nothing is overturned yet — but it sets up JUNO's much larger detector, due to report its own flux this year, to land on an already-diversified map instead of a lone reading.
[2026-08-09] · src arXiv:2510.21976 (DESI Collaboration)
DESI DR2's Lyman-alpha forest BAO measurement (from roughly 2 million quasar sightlines), combined with CMB and galaxy BAO, prefers a time-varying dark-energy equation of state (w0waCDM) over flat LambdaCDM at about 3.1 sigma. Adding different Type Ia supernova compilations swings that number between 2.8 sigma (Pantheon+) and 4.2 sigma (DES-SN5YR); Lyman-alpha plus galaxy BAO alone, with no CMB, gives only 1.6 sigma.
This is a new, largely independent tracer added to DESI's dark-energy program, and it pulls in the same direction as the galaxy BAO results from DR1 and DR2 rather than washing the earlier hint out — which matters, because a systematic peculiar to one tracer would more likely have pulled the two apart than reinforced each other. But the significance still moves by nearly a factor of two depending on which supernova catalog is bolted on, which is a symptom of a result that isn't yet standing on its own. Read honestly: the cosmological-constant-versus-evolving-dark-energy question is not resolved here, only kept alive at a slightly firmer, still inconclusive level.
[2026-08-09] · src arXiv:2511.08493 (Google Quantum AI + Google DeepMind); Nature (2026), doi:10.1038/s41586-026-10759-2
On the same ~101-physical-qubit, distance-7 surface-code memory that first went below the error-correction threshold in 2024 (0.143% logical error per cycle), Google now routes the code's own syndrome data to a reinforcement- learning agent that continuously retunes physical control parameters (gate timings and frequencies) during the computation, instead of using that data only to decode logical errors after the fact. Combined with a newer decoder (AlphaQubit2), the logical error rate drops to 7.72(9)x10-4 per cycle — roughly half the 2024 figure — and logical-memory stability against deliberately injected parameter drift improves 3.5-fold. Numerical scaling tests on simulated codes with tens of thousands of control parameters show the RL optimization cost holding flat with system size. This is a scaling and stability improvement on an existing processor generation, not a new quantum-advantage claim or a bigger code.
Below-threshold error correction proved fault tolerance was possible in principle; it says nothing about whether a processor can hold that performance as hardware drifts over the course of a real computation, which gets worse, not better, as qubit counts grow. This is the first demonstration of that self-calibration loop running at the scale of a full error-corrected processor rather than on isolated gates, and the fact that the RL agent's cost doesn't scale with the number of control parameters in simulation is the more consequential number here than the error-rate improvement itself — it's a claim about whether this approach survives being made bigger.
1
[2026-08-09] · src build/orrery.el
build/orrery.el line 1: ;;; orrery.el --- Build entry point -*- lexical-binding: t; -*-.
Every file under build/ carries the same cookie. Elisp was dynamically
scoped for its first three decades; lexical binding is a per-*file* opt-in,
read from that comment before the file is loaded or compiled, not a global
setting you flip once.
The difference shows up exactly at lambda. With the cookie on, a lambda
captures the bindings visible where it is written — it becomes a real
closure, carrying its lexical environment with it wherever it is later
called. Without the cookie, a free variable inside a lambda is looked up
dynamically, at call time, by walking the stack of currently-active
bindings; if none are active, it falls through to that symbol's global
value, or errors if it has none. The lambda expression itself is just a
list — it captures nothing.
Search build/ for lambda and you find plenty — orrery-render-html.el:335
(seq-map-indexed), :338 (a sort predicate), orrery-feed.el:105
(another sort predicate), orrery.el:67 itself. None of them would actually
break if the cookie vanished, and that is the trap, not the reassurance:
every one of those lambdas is called immediately, inside the same dynamic
extent as the let or parameter that defines its free variables, so dynamic
lookup happens to find the right binding by luck of timing, not by design.
The cookie is what makes that stop being luck. The day someone stores one of
these lambdas in a plist and calls it later — outside the original let —
dynamic scoping breaks silently or errors on a variable that "should" still
be there, and the bug reads like a mystery until you remember the file
didn't ask for closures in the first place.
Exercise (verifiable): save this as /tmp/closure-test.el with no cookie:
(defun orrery-lesson-adder (n)
(lambda (x) (+ x n)))
(princ (funcall (orrery-lesson-adder 10) 5))Run emacs -Q --batch -l /tmp/closure-test.el. It signals a void-variable
error on n — by the time funcall runs the returned lambda, the parameter
binding of n from orrery-lesson-adder's own call has already gone out of
dynamic scope, and n has no global value to fall back on. Now insert
;; -*- lexical-binding: t; -*- as the file's first line and rerun: it
prints 15. Same code, same call sequence — the only change is whether that
lambda closed over n or not.
2
[2026-08-09] · src twelve-keys/src/exercises/upper_structures/exercise.py
Key: F♯7 (V of B) — six sharps, and the least-visited key in most players' V7 vocabulary. A real test of whether the upper-structure shapes are actually memorized, rather than just pattern-recognized in the easy keys.
5 minutes: right-hand major triads only, root position, ascending chromatically from C through B and back down. No left hand, no chord context — isolate the triad shape by ear and hand before asking it to sit over a dominant.
Build both upper-structure triads for F♯7. LH holds F♯ + E (root + ♭7, the tritone that spells "V7"). RH stacks, in turn:
♭V/V (altered) — C major (C–E–G), a tritone above the root: gives ♭5, ♭7, ♭9.
II/V (Lydian dominant) — G♯ major (G♯–B♯–D♯), a whole step above the root: gives 9, ♯11, 13.
Success criterion: both triads land cleanly and in tempo at 72 bpm (quarter-note pulse), cycling ♭V → II → ♭V → II over the held F♯7 shell, with no pause to find the shape.
A ii–V–I in B major. Voice the V bar with the altered UST (C major) on beats 1–2 and the Lydian-dominant UST (G♯ major) on beats 3–4 — moving between the two within one bar, the way the exercise notes describe Mehldau doing it — then resolve to a plain B major shell on beat 1 of the I. Shape it so the ear hears the tension release, not just two chords going by.
Stop when the triad reach starts costing time — any hesitation before the RH lands, or the LH shell drifting to compensate. That means the shape isn't automatic yet, and drilling past that point trains the hesitation, not the fix.
[2026-08-09] · src —
hearBrad Mehldau — All the Things You Are (The Art of the Trio, Vol. 4: Back at the Vanguard) · Tidal · Spotify
The upper_structures exercise notes name this tune directly as the one to test the ♭V→II pair on — its V chords sit still long enough to hear the upper voicing brighten right at the resolution. Listen for that moment against today's technical target.