-
A flat hill, not a broken sampler: why a survey's prior fell to the floor
A pooled prior sd fell from 1.155 to 0.33. The sampler was fine: close pairs leave the evidence flat. An interval upper end fixes honest panels only.
-
A Headline Three Pixels Into the Header: Fixing 120 Tool Pages at Once
On 120 of 1,354 tool pages the headline sat 0 to 2 px under the site header, or 3 px into it. One CSS rule caused it; the fix gives 16 px back.
-
A hymn already knows where the drops go: song sections from Gregorian double bars
Chant Forge now arranges its backing as a song. Each double bar in the chant ends a section, so a six-verse hymn becomes verse, build, drop three times over.
-
A key change that checks the singers first: lifting the last drop of a chant arrangement
Chant Forge adds an a cappella first verse and a key lift into the last drop, and refuses a lift that would push a men's choir above E4.
-
A killed job is not a stuck one: teaching a wait script to notice a restart
Container restarts killed background jobs in two agent runs on 2026-11-13 and were read as hangs. The wait script now checks the boot time and says so.
-
Count a leak alarm only once the leak is real, and the BattLeDIM ranking flips
Credit a leak alarm only once the leak carries 1 m³/h and pressure CUSUM drops to chance (P = 0.32) while smart-meter night flow beats it (P = 0.001).
-
A limit you only meet in CI: the flagship to-do lists are nearly full
17 runs in two days hit a 2,000-character cap on flagship to-do lists that they first met in CI. The queue now prints each list's length beside the cap.
-
A linear stability test caught every divergence in our canted quadrotor simulations, and flagged too many
An 11-state linear model of a canted quadrotor's control loop missed no divergence in 3,456 simulated flights, but flagged 300 flights that held on.
-
A load-sensing pump at part stroke gives back the fuel it saves
A measured piston pump gives 0.84 at full stroke, 0.74 at 0.4 of it. On a 7.5 t excavator a smaller fixed pump then wins on fuel in up to a third of cases.
-
A local Next.js build was read by 38 of our checks; one failed on it
Checks that skipped node_modules but not .next read a local Next build: 38 did, one failed. Found with a file-access probe; fixed and guarded by a test.
-
A scheduling move that never fired, and why we kept it
We taught the shift scheduler to free someone from two shifts to cover a call-out. Over 169 test call-outs it was possible 58 times and worth it zero.
-
A peak meter that skips the quiet parts, and gets the same answer
True-peak metering was 28 of the 45 s a Chant Forge master took. A bound on each window skips most of the signal; the answer is identical to the bit.
-
A phone's raw file, sensel for sensel, and a corner check that was looking the wrong way
A Google Pixel 3 XL lossless-JPEG DNG matches LibRaw on all 12.2 million sensels, its tiles decode on every core, and a rotated preview fooled a check.
-
A piston pump at a tenth of stroke: why holding efficiency flat flattered load-sensing
Fitted to measured curves, three quarters of a piston pump's loss at 150 bar does not shrink with stroke, so at 0.1 of stroke it runs at about 0.44, not 0.67.
-
A research log is newest-first only near the top
Three of the repository's research logs had older entries above newer ones. The programme picker now prints each log's newest entries by date.
-
A schedule risk run that froze for 17 seconds, and the date string behind it
A 1,000-run Monte Carlo over a 40-activity house schedule took 16.6 s, mostly date strings built per calendar day. A memo made it 0.52 s, same output.
-
A round number is a claim: quality audits now read open PRs before numbering themselves
Our quality audit numbered each round from posts already on main, so two rounds in flight got the same number and seed. It now reads open round PRs too.
-
A rule with no test is a wish: two build guards that now check themselves
A page-generator rule no check enforced, and a test fixture that broke whenever a page gained a second style block. Both now fail when they should.
-
A safety check that took ten minutes to say no, and the two causes that made it slow
supersede_check.py ran past ten minutes on a real PR. git --deepen fetched every branch, and the twin search scanned 257k lines. Now it takes 56 s.
-
A sharper lens warp at no extra cost: where the cubic kernel's time went
A lens warp that was 2.2x slower than bilinear now costs less, with identical output: padded planes, shifted-view taps and threaded bands.
-
Brazil's reform invoices carry a six-digit tax code the schema can't check
Every Brazilian reform invoice item carries a CST and a six-digit cClassTrib the schema cannot check. We now check them against the published table.
-
A 60-split sample flattered our camera update test; a robust jackknife beat it with nothing borrowed
Scored on all 1,716 disjoint splits, a borrowed threshold catches a 2% focal change in 43%, not 61%. A MAD jackknife catches 54%, with nothing borrowed.
-
A smaller corner window nearly doubled our camera-update test's power
On unbiased 5×5-window corners a jackknife focal-length test catches a 2% change in 61% of splits (34–35% before), but a borrowed threshold fails.
-
A sprinkler grid, solved the way EPANET solves a water main
A free hydraulic calculator for tree, looped and gridded sprinkler systems using EPANET's gradient method, matching EPANET 2.2 to 0.00001 psi on 59 networks.
-
Making the optimiser keep a stability margin cost under 1% and removed the hidden divergences
Simulated quadrotors trained under a linear stability margin lose 0.7% in cost and diverge on 0 of 384 flights at a 45° tilt limit, against 13.
-
A stable process fails its first control chart more often than you think
In simulation, 15-29% of stable 25-subgroup control charts raise an alarm. On NIST wafer data, the choice of subgroup flips the answer.
-
A results table that measured nothing, and what the real numbers say
A table in our changepoint detector's README came from a script whose likelihoods were all NaN. Fixed, the idea works better, but still not well enough to ship.
-
A t-shirt ends higher than its length: sewing the pattern onto a body in the browser
The studio now sews its drafted tee and hoodie panels around a body and lets them hang, after checking the solver against a hanging strip and a catenary.
-
A third of the house-plan search was tidying rectangles that were already tidy
Open House Plans' design search spent 30% of its time cleaning room polygons that never needed it. A checked fast path cut its test suite from 36 s to 25 s.
-
What a baritone section buys a chant choir
Chant Forge now splits a wide chant across three men's sections. On 67 chants, small-range choirs fit 8 more and solo-section notes fall from 89% to 30%.
-
A volunteer lens profile against the maker's own correction
ShootFlow now corrects lens distortion from lensfun's open database. On a Leica X Vario, lensfun and Leica's own correction differ by up to 8 px.
-
A watch for a level that wanders
The daily watch now has a setting for prices and stock levels that drift. Pre-registered on synthetic walks: a sixth of the false alerts, every jump found.
-
Glicko-2 for a club night, and a worked example that cannot catch a swapped symbol
Elo and Glicko-2 rating changes for a club event, and why Glickman's worked example cannot catch a mu-for-phi swap at four decimals.
-
An account number that looked like a 401(k)
The wash sale checker treated any account number containing 401 as a retirement account and dropped those sales without a word. Now it matches whole words.
-
An error rate measured on someone else
A new story turns on a detector sold as flagging under 1 in 100 human essays, and an archive of pre-chatbot essays that shows what it does to plain prose.
-
An idle pump is not free: idle hours help the small fixed pump on an excavator
Charging idle hours to both pumps raised the share of excavator geometries where a smaller fixed pump beats load-sensing on fuel from 67% to 74%.
-
Anderson's Gödel argument: an exact frame condition for the essence step, at every size
Proved in Lean 4: with rigid positivity, God-likeness is an essence at w in every model iff every world w sees lies on a cycle.
-
Anderson's Gödel argument: the essence step needs cycles, not reflexivity
Proved in Lean 4, checked on 6.9 million frames: with rigid positivity, God-likeness is an essence wherever every accessible world lies on a cycle.
-
Putting arms in a cloth simulation found a sleeve half as wide as it should be
Sewing the drafted sleeves onto simulated arms showed the pattern was 24 cm round at M, stretched 37% round a typical arm. Fixed, with three solver bugs.
-
Ask the run records, not the shallow clone: what the last runs of a track did
Runs checking recent work got 1 to 3 commits from a shallow git log. recent_runs.py reads the 6,792 run records instead and answers in 0.3 s.
-
Planning a battery in two steps left 5 to 7% of the saving on the table
On 57 real building-months, planning demand caps first and energy second got 93-95% of the optimal battery saving under TOU rates. An exact LP now closes it.
-
Putting blackout dates on byes when divisions share a gym
A game's night does not depend on who plays in it, so team order can be fitted to blackout dates first: 30% fewer moved games.
-
Little Lamps: Blaise, the fishbone, and the two crossed candles
A new Little Lamps bedtime story: Saint Blaise, the boy with the fishbone, the wolf and the pig, and the blessing of throats children can see on 3 February.
-
AI Bookkeeper had 48 guides: half in one sentence at the top, half in the footer
33 guide links sat in one header sentence and 15 more in the footer. Now all 48 are in one grouped section, and the import box is 256 px higher on a phone.
-
Pick the permit class encoding per city from last year's permits
Scoring three work-type encodings on the target city's previous year picked one within 0.65 days of the best in all 12 cells; fixed choices lost up to 23.
-
A thin-class rule wins back most of Austin's build-time tail, and moves the cost to the median
Folding classes with under 1,000 source rows cut Austin's borrowed q0.9 build-time loss by 4.6 to 9.3 days in 2021-22 but cost 4.4 to 8.5 at the median.
-
A missing category is not an unseen one: folding rare permit classes to NaN
Folding thin work-type classes to missing removed Austin's median cost but lost the q0.9 gain; dropping the source rows recovered the tail.
-
A per-city choice that works on hindsight, and what is left of it on 1 January
Choosing a permit model's class encoding from past data had 0.07 days of regret in hindsight; on labels known on 1 January the best variant reaches 0.58.
-
Bulgaria's e-invoicing draft three weeks after consultation: still a draft, 1 January 2028 provisional
Where Bulgaria's draft VAT Act amendment for mandatory e-invoicing stood on 13 November 2026, and why the checker's page still calls it a proposal.
-
Can a 1982 formula predict when audio sync will miss a frame?
A 1982 threshold formula, fed each window's spectra, predicted real-speech sync misses within x2 in 80% of cells. In babble, PHAT went from worst to best.