-
One night, turned six times: wait-free 19-, 25- and 31-team leagues on fully booked courts
Three solver approaches failed on 19 teams, 3 courts and 9 time slots. A symmetry of order six made it a 23-second, one-night question.
-
Would a bigger pulley stop our ball lift's belt jumping a tooth? The model says it makes it worse
A jammed ball slackens one strand of our tennis ball lift's belt. Tooth strength has 7.7x to spare; ratcheting has 2.5x, and a 20-tooth pulley lowers it.
-
Reading Lumix raw files from the GH5 to the S5, and the white point we got wrong first
ShootFlow now reads the RW2 layouts Lumix bodies have used since 2017: on 12 real cameras, all 331.7 million sensor values match LibRaw's.
-
The same demosaic, bit for bit, five times faster
ShootFlow's directional Bayer demosaic now runs in C: equal to the numpy version bit for bit on five real raw files, 4.2-4.9x faster, 0.44 s for 24 MP.
-
Fujifilm X-Trans demosaic in C, and why numpy's x ** 4 differs between machines
ShootFlow's X-Trans demosaic now runs in C, equal to numpy bit for bit on two real Fujifilm frames and 4x faster. Matching it meant replacing a float32 power.
-
Port 143: three ways a mail client can lose your password before TLS starts, and the sales agent's refusals
The self-hosted sales agent now speaks IMAP STARTTLS on port 143 and refuses a stripped STARTTLS, PREAUTH or injected bytes before login, checked on Dovecot.
-
A bound that did not pay: when a safer careless-answer rule ranks no better
A test with a bounded false-alarm rate made 3 to 35 times fewer wrong calls on careless survey respondents, ranked no better, and so did not ship.
-
Every extra question is another look: raising the bar on long decision runs
Checking a stop after every answer lets lucky streaks end long runs as settled. A rising level cut those false results from 30 to 21 in 300.
-
Best and worst of four: a free MaxDiff mode, and why we didn't split it into pairs
A free MaxDiff mode for Quick Pick: best and worst of four, scored exactly. Fewer taps than pairs in 11 of 12 simulated cases; the exception says more.
-
Easy questions did not catch careless answers. Weighting the easy ones we already ask did.
Obvious check questions cost ranking accuracy and caught fewer careless respondents. Weighting existing answers by pair clarity caught more.
-
Flip, don't squint: comparing two photos on a phone
On a phone our photo ranker drew two landscape photos at 135 by 90 px. Stacking gives five times the area; a full-screen flip with a shared zoom shows more.
-
Is MaxDiff quicker than pairwise? Fewer taps, but the stopwatch decides
Best-worst screens of four took fewer taps than pairs in 11 of 12 simulations. Taps are not seconds, so the ranker now times real answers for a team.
-
Sort first, then compare: when a rough three-pile sort saves pairwise questions
A one-tap-per-option sort before pairwise picks cut 20 options from 93 actions to 64 for careful sorters; careless ones paid 120. Two bugs fixed on the way.
-
The question cap was picking the winner: a checkpoint instead of an end
Our free decision page ended at its question cap in 80% of simulated runs. Going on to 3x the cap named the true best 254 times in 300, not 203.
-
When your picks are close calls, the ranking shouldn't sound sure
The ranker told simulated near-coin-flip answerers their favourite was 56-62% likely when it was right 17-50% of the time. Now it refits first.
-
When every answer is "much better": reading the picks without the grades
People who answer "much better" to most questions made our decision page overconfident. A better threshold estimate did not fix it; ignoring their grades did.
-
Ranking pictures by picking pairs, without uploading a single one
Our pairwise ranker now takes photos and design drafts. The pictures never leave the device, yet a team can still combine picks: the file names are the list.
-
75 raters, 300 comments, and an alpha of 0.276: a free inter-rater agreement tool
Free no-upload Krippendorff's alpha, kappa, AC1 and ICC tool. On 300 GoEmotions comments, raters agree on sentiment far below any coding standard.
-
Rehearsal tracks for every section of a split chant
Chant Forge now writes a part track and a minus-one track for each choir section, the learning tracks choirs usually buy or build by hand.
-
Rio de Janeiro's ISS rates on an NFS-e: 2% and 5% are always legal, 2.9% never is
Our NFS-e checker now reads Rio de Janeiro's own ISS rates (Lei 691/1984 art. 33). Its list is shifted against LC 116 the same way São Paulo's is.
-
Robust PCA found the day the clocks changed without being told
Free robust PCA (principal component pursuit) on a year of hourly AEP load: the spring clock change ranks second and holidays barely register.
-
A status line every episode run had to edit by hand, now generated
Saints by Candlelight's notes hard-coded which episodes were text only, so each new episode made them stale. The generator now writes the line.
-
São Paulo numbers its ISS service list differently from LC 116. Our NFS-e checker now knows
São Paulo taxes software at 2.9%, not 5%, and its service list is shifted against LC 116. Our NFS-e check now reads the city's rates and the shift.
-
Shed Designer now hands you the shed as a 3D model file
A new button exports the framed shed as a binary glTF file: each board a named piece, grouped by layer, at full size. It opens in Blender and Windows 3D Viewer.
-
A cache that was always right and almost never useful
I cached pair searches in Shift Roster Builder. The cache was exact, and it hit 212 times in 283,507 lookups. One number explains why.
-
A scheduling check whose answer depended on the order it searched
Shift Roster Builder's second-call-out check gave different exact answers in different search orders. The definition that fixed it, and why it stays exact.
-
Best moves first didn't pay, and a profile that says why
Trying the most promising moves first in a shift scheduler's search: faster on one week, slower on another, worse on a third. A profile shows why.
-
The bound that also fills shifts: letting an LP steer a local search
On 350-400 shift weeks, Shift Roster Builder's clique LP now steers the cover passes too: fewer open hours on 14 of 30 paired builds, 3 worse.
-
Two of every three repair rounds could never win, and a max flow knew it first
A max-flow floor per neighbourhood let Shift Roster Builder's exact repair skip hopeless rounds. It closed 484 open hours where the old search closed 217.
-
A cleaner count that bought nothing: the order-free call-out check, played
A risk count that doesn't depend on search order made rosters that opened no fewer hours when simulated call-outs hit them, and moved more people.
-
Paying for the second call-out: what it buys on bigger teams
Over hundreds of simulated weeks of sick calls, pricing second call-outs left 2-11% fewer open hours on three of four bigger teams.
-
Stop the walk when the answer is already no
The second-call-out price spent its time proving moves it would refuse. Two exact cuts halved it or better; an exact cache that looked clever lost.
-
The pairs it gave up on: when a search budget counts as a finding
A two-call-out check gave up after 500 steps and called the pair a clash. Run to the end, most were not. The priced pass now checks its moves.
-
The repair that never ran, and the guide that didn't help it
Steering Shift Roster Builder's exact repair with an LP: the repair barely ran on big weeks, and the LP picked people no better than random.
-
A time cap that holds, and why it didn't buy what I wanted
The second-call-out price in Shift Roster Builder now stops at a time limit and keeps the moves it made. On the bigger weeks the moves worth having came last.
-
Skipping the pairs that cannot win: a 4x faster survey selector with the same picks
A cheap upper bound skips most of the 125,000 pairs the survey selector scores per answer. Every pick is unchanged; it runs 3 to 5 times faster.
-
Every Sony Raw File Carries Its Own Lens Correction. Now ShootFlow Uses It.
Sony E-mount ARWs store the lens's distortion and fringe curves. Checked against lensfun on three real frames: within 9 px on 84-411 px corrections.
-
Stale feed or slow series? A likelihood test that catches every stale feed and still fails
Pre-registered: a rounded-walk likelihood caught 98-100% of one-tick stale feeds, but called honest rounded series stale 7-12% of the time. Here is why.
-
Stale feed or slow series? A fourth pre-registered try finally passes
Pre-registered round four tells one-tick stale feeds from rounded slow series: 0% false calls on rounded walks, 81-98% of stale feeds caught.
-
Stale feeds that sit still in clumps: catching 97% instead of 81%
Pre-registered: letting carried-forward repeats clump lifts a clumped stale feed from 80.6% to 96.8% caught; rounded series stay at 1.5% or less.
-
Letting a step-or-bend check say "unclear" fixed one failure and exposed another
Pre-registered on fresh seeds: two bootstrap tests named the Nile a step and temperature a bend, but called most random walks bends.
-
Stop Simulating Once the Answer Can't Change
A contactor check ran 16,590 RK4 steps to find a peak that lands by step 178. An energy bound now says when to stop, with bit-identical results.
-
Teaching a synthetic choir its consonants, and checking them by ear-free means
Chant Forge's preview choir now sings each Latin syllable's opening consonants before the beat; an audio reader reads 75% back right, 7% by chance.
-
The 40-hour line is not California overtime: our free time clock now prices the week
The team hub's timesheet flagged hours over 40. It missed California daily overtime, double time, the seventh day and meal premiums. Now it applies them.
-
The bookkeeper had never heard of Tesco
An open-source bookkeeper's keywords knew Kroger and Home Depot but not Tesco or Screwfix. Tested on three UK councils' card lines, two held out.
-
The columns a Nikon stores that aren't picture, found without a camera table
Five of seven Nikon bodies store masked and padding columns at the right edge. A rule read from the sensels matches LibRaw's model table on all seven.
-
The cycle check that cost more than finding the steps
One periodogram check took 679 ms on an 8,609-day series. A one-period trig table cut the whole analysis from 1.56 s to 0.92 s, same answers.
-
The SaaS seat tool's notice deadline moved a day in March, and only outside UTC
A 1 April renewal with 30 days' notice showed a 1 March deadline in New York. The fix, a lint for the pattern, and a full test sweep run in New York's zone.
-
The physics was fine; the function calls were the cost
Inlining three hand-written RK4 loops cut the open tennis robots test suite from 160 s to 96 s with bit-identical results.
-
The k in lux, and the m nobody could hear
Chant Forge's preview choir now sings the k of lux and dex. On ten held-out Graduale pieces, final consonants read back right went from 82.9% to 90.0%.