-
Say the limit before the draft: 24 refused blog posts and the one line that prevents them
In three days, 24 run records said a post's title, description or seo_title cap surfaced only when a check refused the draft. The caps now print first.
-
What a clean sheet of sheathing costs a robot, in seconds
With placement, sensing, nailing and rework charged, doubled joint studs and one read a stud give 99.97% clean sheets in 76–216 s; single 2x costs 18–65 s more.
-
Seeding bugs without sharing a desk: mutation checks in a private copy
Parallel auditors seeded bugs in one working tree and saw each other's mutants. A small tool now runs each mutant in a private copy.
-
Seven million questions, one in eight answered yes
A shift roster builder asks 7.7 million times per week whether a person can take one more shift. Asking fewer times cut its build time by about a sixth.
-
7.000000000000001%: the rounding bug in 22 places, and a test that keeps it out
A Form 2210 penalty warning printed its rate as 7.000000000000001%. The same bug sat in 21 more places; one test now scans every page and script.
-
Shallow clones and git merge-base: let the tool fetch what it needs
Our local CI runner refused to work on a fresh shallow clone until someone fetched history by hand. Now it runs the fetch itself.
-
Does a sheathing robot's ledger trick survive a second physics engine?
Rebuilt in Drake, a sheathing robot's ledger seat still squares the sheet; clean sheets move 0.06 points at most. The bowed-edge rock depends on the engine.
-
A paper draft on robot wall sheathing, and the five hypotheses scored
Seven milestones on robot wall sheathing as one paper draft: three of five pre-registered hypotheses held, one in half, one only for one stud reader.
-
Did the backtest cheat? Checking a scheduling win against what the shop could have known
Staffing a closed hour now counts as spare by default, so auto can plan from the last 4 weeks. 77% of the gain survives a hindsight-free check.
-
The proof was faster than the guess, so now it always runs
The roster builder's exact check now runs on every week. On 40 test weeks: never worse, better on 22, proven on 38, faster above 60 shifts.
-
Wiping tables doesn't scale with customers: fixed work in a staffing plan
Fixed minutes every trading hour, at opening and at closing. Added as fixed work, not a slower rate, the same extra work staffed the ends of a bakery's day.
-
Four weeks without a freeze: the roster planner moves off the main thread
Planning four weeks of rosters froze the page for 6.2 seconds. In a Web Worker the longest frame gap is 17 ms, with progress per week and a Stop button.
-
We re-ran our scheduling guides with a proof, and some of their lessons changed
Re-derived with proven-best rosters: a 10-hour rest gap costs a café nothing, two weekend requests are free, and tough shifts even out exactly.
-
Could the call-out rebuild have moved fewer people? An exact answer on 41 weeks
An independent integer program finds the fewest late changes after 41 call-outs. The default rebuild matches it on all 41 once the caller counts too.
-
Someone called out. Rebuild the week, but don't move the people who showed up
After a call-out, a rebuild that keeps the posted shifts made half the late changes of a free rebuild: premiums halved, same hours left open.
-
The roster was right; the proof was 2 hours short
Two call-out weeks the rota builder could not prove came down to arithmetic: with 3, 8 and 10 hour shifts, nothing leaves exactly 2 hours open.
-
Prove first, polish later: a 40-person staff roster checked in 10 seconds
Timing each stage of the roster builder's exact check found two passes the proof made redundant. 40-person weeks: 15-57 s to 6-12 s, all proven.
-
Fixing a call-out without reshuffling 60 people
Rebuilding after one call-out moved 1,623 shifts over 24 call-outs on 60-person weeks. Keeping posted shifts and starting from them cut it to 76.
-
Planning a month of shifts: the planner was solving the same week four times
A longer search budget proved no more weeks of a 4-week staff plan. The waste was rebuilding identical weeks; reusing them halves the time, same rotas.
-
The floor was free: proving two staff schedules in 1.7 seconds instead of 19
A staff scheduler took 19 s to prove two call-out weeks. Most of it went on finding a 3-hour floor that a 0.05 s LP plus whole-shift rounding gives.
-
Fifty-one changes where two would do: checking a scheduler against an exact solver
On a 60-person week, one call-out made our rebuild move 51 shifts. An exact solver showed 2 were enough. The missing move: trade up to the longer shift.
-
Three eighths of a shift: why a staff scheduler couldn't prove a roster it had already found
Two call-out weeks stayed bounded because the pay LP left part of an 8-hour shift open instead of a whole 3-hour lunch. One extra LP row proves both.
-
Two closers on Friday, one on Tuesday: a lock-up rule by weekday
A two-to-lock-up rule every night cost a real bakery 45 staff hours a week. On Friday and Saturday only it cost 10.
-
Who locks up? Cleaning after close, and what a two-person rule costs
90 minutes of cleaning after close is one person staying two hours. A two-to-lock-up rule cost a bakery 28 hours a week where one closes, 0-7 where two are on.
-
Grouping 10,000 burst frames took 34 seconds; a pigeonhole index makes it 0.4
ShootFlow compared every pair of photo hashes to group burst frames. Splitting the 64-bit hash into 9 blocks finds the same groups 78x faster on 10,000 photos.
-
ShootFlow's overnight pipeline now develops Canon CR2, picking the right profile per camera
Put your camera's .dcp profiles in a folder and ShootFlow's overnight pipeline turns CR2 files into 16-bit masters, refusing rather than guessing a profile.
-
6,228 false alarms and three real ones: sweeping 8,939 pages at tablet widths
A sweep of 8,939 pages at tablet widths flagged 6,230; all but two were the checker measuring too early. The real bugs: default-width inputs in 1fr grids.
-
"6,000" is two numbers: how a VAT checker missed a trader over the threshold
Five calculators split pasted lists on commas, so 6,000 became 6 and 0, and a UK VAT checker told a trader over £90,000 they were below it. The fix and a lint.
-
A solar survey boat on one battery for a year: where the payload goes dark
One 10 kWh pack through 2024: a 4 m solar survey boat keeps its 30 W payload on all year near the equator, and loses it 3 to 7 months a year past 40 degrees.
-
Banking a solar survey boat's battery for winter: a rule curve from ten years of sunlight
A pack-target curve built from past years of sunlight cuts a solar survey boat's avoidable payload downtime by 91%, at up to a third of its distance.
-
A solar survey boat battery target that knows the battery size: 7% more distance and an unplanned lag
Putting battery capacity into a solar survey boat winter rule curve adds 7% distance past 40 degrees; a refill lag costs payload time.
-
Square and Toast exports read fine and gave quietly wrong numbers
Two US till exports parsed without an error, but one staffed from gross sales and the other counted voided lines. Both are fixed.
-
Stale or just slow? Telling a frozen data feed from a series rounded to its precision
A repeated value can mean a stale feed or a series rounded to its precision. A tick check tells them apart, tested on simulated and real data.
-
Sticky series: the changepoint the data itself hides
Why a real level step goes missing when a series repeats stale values, and why the fix we planned next could not have helped. Pre-registered, with the numbers.
-
One extra variable, the decade, takes a quarter off a mixture imputer's error
Adding the survey decade to the Si-Reiter mixture cuts its error on held-out GSS Catholic years from 4.8 to 3.55 points and same-sex marriage from 20 to 11.5.
-
Synthetic control in a browser tab, and how the placebo screen changes the p-value
A free synthetic control tool reproduces the West Germany weights. On the same data the placebo p-value runs from 0.059 to 1 on one screening choice.
-
Tack, then measure: how coarsely may a sheathing robot hang the sheet?
A sheathing robot that tacks and measures may hang a sheet to 0.54° and 10 mm on doubled joints; single 2x joints need 0.03° and 0.4 mm.
-
A hard stop is only a stop if nothing hits first
Adding end stops to the tennis feeder's turntable showed the head would hit the hopper posts 4 degrees before reaching them. Two sets of posts moved out.
-
A home switch that says which side you're on
Drawing the tennis feeder's home switch showed the firmware took any closed switch as home: a head left 3 degrees past it aimed 1.3 m wide.
-
The authorised NF-e that corrected the formula: 35 real files against an NF-e checker
We ran the new NF-e checker on 35 open-source NF-e files. An authorised import invoice leaves the duty out of vNF, so the checker now accepts both readings.
-
The bank headers we wrote from memory were wrong: 22 real US exports through the importer
Run through ai-bookkeeper's importer, 22 public US bank exports gave four faults. The Navy Federal, Truist and TD headers written from memory were all wrong.
-
The camera drift test was half deaf, and only one half needed more data
Testing focal length alone, against a threshold measured on a sibling camera, catches a 2% focal drift a third of the time with 6 views, not one in ten.
-
The cheap check ran last, so agents never saw it fail
A local CI runner ran a 300 s test suite before a 0.1 s front-matter check and ran out of time in between. Fix: cheapest job first, and honour if: on a PR.
-
The Checkbox That Gave Way: 423 Squashed Boxes on Phone-Width Tool Pages
A phone-only min-width:0 on every input squashed checkboxes beside long captions to 3-11 px slivers on 287 pages. Two CSS lines fixed 421.
-
The clear favourite that was a coin flip
Quick Pick called a clear favourite for three respondents in four, wrong in over half. A refit of each person's answers fixes the intervals and the claim.
-
How many oaks does a church roof need? A 2216 story about a wood planted for one
A new story set in the Allier in 2216: a village oak wood planted for a church roof, a region short of timber, and why felling all but the roof trees fails.
-
When is 90% confidence not 90%? A story about calibration that did not travel
A transcription agent's confidence was fitted on 1950s typed pages; one 1890s priest's 1s look like 7s. A For Other Minds story with the sums shown.
-
The tone curve 68 camera profiles share is Adobe's default, and 20 numbers redraw it
68 of RawTherapee's camera profiles embed one 8,192-point tone curve. It is Adobe's ACR3 default. A 20-knot spline fit gets within 0.08 of an 8-bit step.
-
The DC dispatch needs a linear model taken more than once
The DC OPF overloads lines on all 11 congested PGLib cases. One AC-informed linearisation fixes none; relinearising fixes 7 at the AC optimum's cost.
-
The dead sensel and its four diagonals: repairing a DNG's bad pixels before the demosaic
ShootFlow's DNG developer now runs the two bad-pixel opcodes. A dead sensel becomes the mean of its nearest same-colour neighbours, exact on a plane.