-
Equally good weeks that are not equally good months
Which of several equally good weeks a staff planner posts decides what next week can do. A one-week look-ahead kept every hour and cut changes 71%.
-
A week of call-outs at a real roster's rates: keep the posted schedule, tell fewer people
At one firm's real call-out rates, rebuilding a posted week from scratch told 6 people a week of a change; keeping the posted shifts told 2.5.
-
Skipping the rebuilds a call-out doesn't need
Two exact short cuts and early stopping cut the call-out check from 419 rebuilds to 243 with the same schedule picked, so it now runs on the page.
-
Charging a manager for slack the team already had
A quick call-out risk check over-counted badly. Counting hand-over chains raised its precision from 53% to 67% and stopped 21 trades that bought nothing.
-
Five times faster on the sample week, and why the multi-week planner did not get the speed-up
Running the exact method before the search made fresh rosters 3.5x faster, same proven score on 80 of 80 weeks. Tied rosters still differ.
-
When nobody can cover, free someone who could
On a 20-person week where no single shift move helped, freeing a stand-in cut the measured call-out exposure from 113 to 68 hours for $35 of score.
-
Reissuing a posted schedule twice as fast, with the same schedule quality
Exact solver first, local search second: 570 posted-schedule reissues after a call-out ran 2.2x faster with the same score, changes and premiums.
-
The 25-hour gap that was never there
A 60-person week sat unproven at 300 open hours over a 275 h bound. The bound was the problem: a 1.3 s LP with clique rows proves the 300.
-
Three ties, not five: a roster counted twice, and a tie-break that changed nothing
Picking among tied schedules at every reissue of a posted week changed 0 of 983 reissues and took 14 times as long. A tie-counting bug turned up too.
-
Six equally good schedules, and the one a call-out hurts least
Equally good schedules differ by half in what one call-out costs. Picking the least exposed cut open hours 3.5% over 40 seeded weeks, with 16% more changes.
-
Should a battery design screen ever skip the cheap model? We tried a switch, and it wasn't worth it
A pre-registered test of a pilot switch between an SPMe trust rule and DFN on every design: it misfired on Ai2020, but the rule never cost more than 12% extra.
-
Six rows, not three: spreading parallel agent sessions over the flagships
40 of 153 flagship PRs on 12-13 November opened within 30 minutes of another on the same product. The session picker now spreads over six rows, not three.
-
Sixteen pixels instead of four: what bilinear sampling costs a lens correction
On a Leica T frame, bilinear sampling in the lens warp threw away 27% of the fine detail. A Catmull-Rom cubic keeps all but 11%, at about twice the cost.
-
Sole trader or limited company in 2026/27: the company never came out ahead
At £50,000 profit a UK sole trader keeps £40,268 and a one-director company £37,994. On a £1,000 grid from £10k to £300k the company never wins. And a bug fix.
-
One calibration budget, several subgroups: the split barely matters, the estimator does
With one budget for several subgroups asked a rotated survey item, Neyman allocation moves the error under 1%. Shrinkage halves it.
-
Splitting our slowest test file did nothing; remembering the last roster did
A 101 s test file looked like the long pole of a 149 s suite. Splitting it changed nothing: the suite was CPU-bound. Remembering repeat builds cut it to 132 s.
-
A battery model's trust flag, without running the model
A closed-form steady state with concentration-dependent diffusivity gives PyBaMM SPMe's minimum salt ratio to within 0.0003 on 408 fresh runs.
-
Stems from a browser tab: a 60-line ZIP writer and a sum check
Chant Forge's studio page now hands a producer phrase markers and per-part WAV stems in one ZIP, byte-equal to the command line, checked in headless Chromium.
-
STL decomposition in a browser tab, and the Mauna Loa CO2 cycle that grew
A free STL tool matches statsmodels to 1e-11. On 52 years of Mauna Loa CO2 it finds the seasonal swing grew about 0.09 ppm a decade, whatever the window.
-
Swap the tuning year and the leak-detector ranking goes
Tuned on 2019 and tested on 2018, BattLeDIM leak detectors lose their ranking: on 32 held-out leaks none beats chance after correction under a flow gate.
-
Teaching a chant scanner to see a diamond
A dependency-free chant scan reader now tells a rhombus from a square (118 of 120 held out), and the 1908 Te Deum was read by eye: 637 notes, no differences.
-
Teaching a scan reader to see a quilisma, and testing it on pages it had never seen
Finding the saw-tooth quilisma in scans of the 1908 Graduale by counting notched rows: set on five chants, then 15 of 17 found on nine never-seen chants.
-
Ten Copies of One Pill: Moving a Repeated Component Into the Shared Stylesheet
Ten tool pages had each copied the same jump-row CSS, and the copies had drifted. One shared rule with custom properties replaces them, with no visible change.
-
Ten cycles of the full battery model rank a hundred
A 10-cycle DFN run ranked lithium-ion electrode designs by 100-cycle energy and found the best design on three fresh cells at a fifth of the cost.
-
Ten posts that broke the site build, and the one-line habit that let them through
A pre-push hook checked post front matter, except on the git add, commit and push chain most runs use. Ten broken posts reached main in four days.
-
The trough that couldn't reach the hopper
Drawing the catch net's ball trough for the open tennis feeder showed balls end up 2.2 m below the hopper. The design note said they would roll in.
-
Stirring a tub of tennis balls without jamming one
The open tennis feeder's tub drains through a 75 mm hole for 67 mm balls, so a few balls can bridge it. A little geometry chose where the agitator's arms go.
-
Tensioning a 2.4 m ball-lift belt: 7.5 mm of travel, and a tension you can count by hand
The top pulley tensioner for an open tennis-ball lift: where 7.5 mm of travel comes from, why 10.5 N, how to set it by counting swings, and a flight fix.
-
72 test suites no pull request ran, and the two that were red
72 of 121 test suites in our CI tooling ran on no pull request. Two were red on main. One script now runs every unwired suite in about 20 seconds.
-
The 1976 warming acceleration, dated and tested: why the test on differences missed it
A level test for a change of trend slope with AR(1) noise and a bootstrap p-value finds the 1976 warming acceleration at p 0.001; differences gave 0.80.
-
The 2024 estimated-tax penalty divides by 366, and our worksheet divided by 365
A red-team run found two gaps from the IRS Form 2210 worksheet: 2024 rate periods divide by 366, and each late payment is rounded on its own.
-
The alert hint on a second labelled set: 47 real server metrics, three pre-registered tests, three misses
Pre-registered on 47 real NAB series watched hourly: the alert hint, a drift lead and a new hint rule all missed. Level detectors sat near chance.
-
The audio cap that outlived its reason: 444 text-only runs and four stories you can hear again
A 400 MB audio cap written for GitHub Pages kept 444 story runs text only after the audio had moved to a CDN. Four new Kokoro readings, and the numbers.
-
We built a fix for the cheap battery model's bias. On fresh designs, the old method never needed it
A gap-aware stop always found the best battery design, but the simpler rule it was built to fix never missed on fresh designs. 788 PyBaMM runs.
-
The full battery model's best fit to a real cell came from two errors cancelling
Against measured Li-ion discharges, SPMe vs DFN moves capacity by at most 1.5 points; the parameter set misses by 4.3 and self-heating matters more.
-
The bearing the model never saw: what 99% on the CWRU dataset measures
On the CWRU bearing fault data, spectrum classifiers score 100% on random and time splits and 55-71% when the test fault is from a bearing never seen.
-
The best battery design sits on the salt flag, so don't throw flagged designs away
A closed-form salt flag saves no time in an SPMe/DFN battery design search, but the best design sat on it in 8 of 21 conditions.
-
The blog's filters now look like one set of controls
The blog's Tag and Sort menus were 28 px tall under a 46 px search box, laid out differently on a phone. They now share its height and one layout.
-
The tennis feeder's CAD tests took 27 minutes because they rebuilt the robot 60 times
The tennis feeder's CAD check ran 19 of 20 CI minutes: an 18 s part build was requested ~60 times. A cache cut the suite from 1637 s to 323 s.
-
The central case hid the pump's drag: a fixed pump now wins in two thirds of excavator geometries
Charging a piston pump's drag below its smallest measured stroke doubled the excavator geometries where a fixed pump beats load-sensing on fuel.
-
The certificate that hid the takeover: why our subdomain checker could almost never say 'confirmed'
Our subdomain takeover checker probed HTTPS only, so dangling S3, Heroku and GitHub Pages CNAMEs failed TLS and could never read as confirmed.
-
Ask the cheap question first: a page check from 5.4 s to 1.2 s
Our missing-element-id check spent most of its time collecting ids on pages that never look one up. Asking that first cut it by three quarters.
-
The choir's changes of singer were not cameos
A three-section chant split tripled changes of singer. A cost per entrance was meant to cut the short ones; on 67 chants there were almost none.
-
The Cold Chain: a vaccine fridge, a week of cloud, and three vials nobody could read
New hopeful science fiction: a week of cloud in Vanuatu, 2049, warms a solar vaccine fridge four days before a measles campaign for 214 children.
-
The colour fringe fits in fifteen terms
ShootFlow now corrects lateral colour from lensfun with distortion, exactly, as one DNG warp. Uncorrected, 30% of lens profiles leave a 1 px fringe.
-
The commit that touched every file: a shallow-clone trap in our router
In a shallow git clone the oldest fetched commit looks like it changed every file. Our router reported its date as 'last landed' for 133 of 1,353 projects.
-
The failed import that cost more than the copy: placing raw tiles on every core
Placing tiles on the decode threads made a Pixel 3 XL raw read 1.3-1.4x faster with numpy. Without numpy, a retried failed import made it slower.
-
Pricing the repair correctly still does not beat doing nothing
Pre-registered on 51 audit cells: with the second-phase floor priced, the bias-free fallback nearly ties the shipped guard but never beats it.
-
The fallback that only fails when it is rare
Pre-registered on 45 audit cells: a guard-first fallback adds 20% width gain but undercovers where it is rare, since the pilot that chose the design is reused.
-
The fastest cutting-stock solve is the one you already did
A framing takeoff re-solved every lumber size's cutting-stock problem on every edit. Keeping 128 recent answers made a test 4x faster, output identical.