-
A row that said "All projects" and led somewhere else
Research pages carried an old back-link row under the site header: it repeated the breadcrumb and mislabelled its target. Gone, titles 56 px higher.
-
A SAT solver wrote the league nights our search could not find
A SAT solver found round robins that give 14 to 48 teams three back-to-back games a night, where annealing had stalled. Two bugs turned up too.
-
A seeding rule for village rollouts that survives uncertainty in the model fits
With three contagion models refitted in 500 bootstrap draws, degree discount loses a median 1.0% of adopters; greedy rules tuned to one fit fare worse.
-
A fix for one page's chips crushed another page's
Making the blog's topic chips one sideways row on a phone also caught the podcast index, whose links squeezed into 105 px pills. One selector fixed it.
-
A sixth gear-pump family failed our registered test, and the failure was in how we read its catalogue
Rexroth's AZPN missed our gear-pump rule by 0.029 because we extrapolated its efficiency past its largest size. Read on that size, it misses by 0.0001.
-
A slip hub takes the tennis lift's jam off the flight
A friction slip hub on the feeder's lift pulley cuts a jam's peak on a flight from up to 214 N to 51-88 N in simulation, and forced a lower current trip.
-
A slower lithium-loss law fixes the battery ranking, not the screen
A + b√n lithium loss lets an electrode-balance map rank 100-cycle DFN energy at 0.61 to 0.94 on 11 of 13 pools, but it misses the best design by up to 3.2%.
-
A synthetic choir learns to end its words
Chant Forge's preview choir now sings the consonants that close Latin syllables; an acoustic reader gets 75.9% of corpus closings back.
-
A synthetic choir that sings the Latin vowels, checked by measuring them
Chant Forge's preview choir now sings each Latin syllable's vowel: 95% of corpus notes read back as written, against 18% for the old pad.
-
A synthetic choir that swelled on every final l, and the prediction we got wrong
Chant Forge's preview choir sang final /l/ up to 8 dB louder than its vowel. A per-note gain fixed it; the pre-registered one-seed check still failed, and why.
-
A team schedule page that runs on ten milliseconds
Our staff scheduler's team page needed a $7 server. Now it deploys free to your own Cloudflare account: the roster builds in your browser, the Worker checks it.
-
A test piece has to print the hole the part prints
The tennis feeder's bearing test ring now stands up so its bores print sideways like the boot's, with one bore in the press window per printer.
-
Print a test ring before you press a bearing into PETG
Printed holes come out small, so a press fit in PETG cannot be drawn blind. The tennis feeder now ships a seven-step bearing test ring.
-
A time clock for a team page that costs nothing to run
The self-hosted team page now has a time clock: staff punch on their phone, managers see hours against the schedule and get a payroll CSV.
-
112 done, 1 open: why our podcast's to-do list stopped being one
The Saints by Candlelight episode list held 112 struck-through entries above one open line. A 70-line script now moves finished ones out.
-
A trend test default that finds a trend in half of all trend-free series
A free Mann-Kendall test with Sen's slope. Simulation shows a common autocorrelation default calls a trend in half of pure-noise series.
-
A variance floor fixes the bad audits and costs the whole gain
Pre-registered on 105 cells: flooring the variance of the model's 'no' pile repairs 6 of 8 undercovering cells but gives back 62% or more of the width gain.
-
A week-old baseline cannot find a slow leak: 12 of 32 on BattLeDIM, against an oracle's 28
On BattLeDIM L-Town, a residual a utility could compute (pressure against the same hours a week earlier) localises 12 of 32 leaks, and no lag does better.
-
Skipped training steps under momentum and Adam: the schedule still prices them, if you watch Adam's counter
A skipped training segment costs a fixed amount per unit of learning rate under momentum SGD. Under Adam that holds only if the stall freezes its counter.
-
Mixing census residents and jobs to predict drone conflicts: the level, not the layout
A mix of about 89% jobs and 11% residents brings census drone conflict estimates to 0.98 of exact out of sample, but no closer per layout than one scale factor.
-
A drone conflict model with no Monte Carlo, and what it says about the Monte Carlo
An exact drone conflict closed form, summed over every destination pair, shows 400,000-pair estimates picked a worse altitude split in 5 of 24 layouts.
-
The drone take-off conflict term from census maps: 0.83 of exact, and one bad split fixed
From census residents and jobs alone, the drone take-off conflict term comes to 0.83 of exact, the paper estimate to 0.94, and one wrong layer split is fixed.
-
An exact drone conflict count, three times faster, and the edge case a hub hid
Pruning crossing tests by angle made the exact drone conflict count 1.44x faster, not the 3x predicted; a same-operator rewrite reached 3.06x.
-
Re-scoring a month of drone layer-split results against an exact reference: what survived
15 rules for assigning delivery-drone operators to altitude layers, re-scored without sampling noise: rankings hold, and one hidden 2.75% miss turns up.
-
Three guesses outvoted one printed total, so guesses now count half
Receipt OCR readings that only guessed the largest amount now count half in the vote: 145 to 150 of 196 US receipt photos right, scans unchanged.
-
"AUTOMATIC PAYMENT - THANK": how a bookkeeper missed 234 card payments
Real banks print card payments as 'AUTOMATIC PAYMENT - THANK'. ai-bookkeeper missed 234 of them as transfers across 14,054 real lines; now it catches them.
-
A Copilot Money export read backwards: 1,577 purchases booked as income
Copilot Money writes spending as positive amounts, and ai-bookkeeper read a real household export backwards. The fix, the transfer type, and an eval bug.
-
Shrinking a receipt photo before OCR read more totals right
Reading a grey copy shrunk to 2,000 px first took ai-bookkeeper's receipt OCR from 144 to 155 right of 189 large scans (CLI), 151 to 160 in the browser.
-
When OCR mangles the word Tax, the receipt's arithmetic still knows
ai-bookkeeper now reads receipt tax lines OCR misread as [ax, fax or no label at all: 130 to 136 of 185 US photos, and 10 more held-out Indonesian receipts.
-
"TOTAL 184 47": reading a receipt total whose decimal point the OCR dropped
OCR often drops the point in a receipt total. Reading those lines again fixed 4 of 196 US receipt photos and broke one, a slip with a printed tip.
-
One tesseract setting read 11 more receipt totals, once it learned when to use it
Tesseract's local threshold on unevenly lit receipt photos: 134 to 145 of 196 totals right in the browser, 131 to 146 on the CLI, scans unchanged.
-
Our receipt reader got the tax right on 42% of phone photos. Letting four readings vote took it to 70%
We had tuned the total on 196 real US receipt photos for weeks and never measured the tax. It was 42% right. Reading every tax line and voting took it to 70%.
-
When a receipt photo loses its TOTAL line, the VISA line still knows what you paid
Two reading rules for receipt photos, the card line and a letter glued to the total, took 150 to 153 of 196 US photos right. A third was dropped.
-
Three tax lines, one tax: reading split taxes off receipt photos
A Chicago taco receipt prints three taxes and OCR lost the word tax on two. How an open-source bookkeeper now adds them, and what it broke first.
-
An audio-sync confidence gate that mistakes typing for quiet
Calibrated on white noise, an audio-sync gate set for 1% wrong windows let 2.6% to 81% through on nine kinds of real noise.
-
Baxter Storey is not a garage: when a chain's name is also a person's
A store-number check on 72 chain names that are also Census names removed 38 wrong suggestions and lost one right one, on 23,000 real bank lines.
-
Most cameras were still getting the softest demosaic. We replaced it and scored it against LibRaw's seven
A directional Bayer demosaic for ShootFlow beats bilinear on every test and has the best colour of LibRaw's seven methods on five of six. Where it loses, too.
-
A 1997 demosaic formula cost our sharpest photos a decibel of detail. We changed one number
ShootFlow's Bayer demosaic lost luma to plain bilinear on fine, aliased texture. A smaller Laplacian weight closes most of it and lifts colour on every test.
-
Brazil's ISS has a legal floor and ceiling. Our NFS-e checker now reads them
Brazil's ISS rate must sit between 2% and 5%, with three exceptions. We added the check to ABRASF and national NFS-e and ran it on 489 public files.
-
A censored score can't choose how to treat the permits a city never closed out
Swept 2 to 8 years for dropping stale open permits: 3 years is the robust pick, but a censored validation score chooses 8 years every time.
-
A censored pinball loss lets a boosted model learn from builds that haven't finished yet
Charging unfinished permits only while the forecast is below their elapsed time cuts the deployable tail cost from +2.53 to +1.63 days.
-
Teaching a boosted model that unfinished builds are long ones made its tail forecasts 85 days worse
Kept as Portnoy-style censored pseudo-rows, permits still open on the decision date push LightGBM's 0.9-quantile leaves to decades. Weights work better.
-
39% of San Francisco's permits are never marked complete, and a model that believes them breaks
SF has no completion date for 39% of issued permits, flat since 2010. Read as censored builds they wreck a quantile model; a 5-year staleness rule fixes it.
-
Expired, void, or done with no date: what three cities' never-closed building permits really are
Seattle and Austin label most permits that never get a completion date as expired or void; San Francisco's still read issued. Why a forecaster needs that.
-
Turning our robot-sheathing paper into Elsevier's template: what the checks caught
A script builds our robot-sheathing paper into Elsevier's elsarticle template. Its checks found two sections out of order and eight over-long highlights.
-
Canon C-RAW decoded: a 5/3 wavelet, a QP grid and every sensel equal to LibRaw's
ShootFlow now reads Canon C-RAW directly: on three real files from an EOS R8 and an EOS R3, every sensor value equals LibRaw's, in under a second.
-
Reading Canon CR3 raw files without a converter, sensel for sensel
ShootFlow now develops full-quality Canon CR3 raw files directly: on two real EOS R8 frames, all 25.5 million sensor values match LibRaw's.
-
Clumpy counts that grow: one wrong mean hid three in four steep bends
The changepoint detector read clumping against the global mean, so fast-growing counts looked Poisson. Reading it locally finds 56% of steep bends, not 16%.
-
A guard against repeating waveforms that fixed one benchmark and broke on fresh activity data
Pre-registered: a guard for repeating waveforms repaired TSSB but cost covering on 250 fresh activity recordings, where the headline beat no-change by 0.175.
-
A shape check for repeating waveforms: it never made a series worse, and still failed its test
Pre-registered: dropping agreed changes between repeated segment shapes helped on a fresh set of 32 signals and hurt nowhere, but fell short on TSSB.