-
A ratchet inside the battery plan: worth 0.8% of the saving, and now exact
Planning a battery month by month under an 80% ratchet left up to $660 a year unclaimed; one whole-year LP gets within $2.29 of HiGHS on 72 building-years.
-
A raw decoder path no camera we have takes: point-transform scans move to C
ShootFlow's DNG reader sent lossless-JPEG scans with a point transform down a slow Python path. Now they decode in C, up to 54x faster, same samples.
-
A sample invoice that went stale in two weeks
An offline check of NF-e presumed credits (cCredPres) against the NT 2025.002 rules found 6 rejections in 15 public files, one a vendor sample gone stale.
-
A semicolon in the memo hid a whole bank statement
A real US Bank card export read 0 of 57 rows: its memo had more semicolons than the line had commas. The fix, and the merchant code the memo carried.
-
Ignore moves smaller than X: a floor per sensor cut false alarms by two thirds, and missed its bar anyway
Pre-registered on SKAB's 33 pump faults: a per-sensor floor read off a normal run cut false alerts from 6.34 to 2.27 per 100 checks, but lost 5 faults.
-
A staff line nineteen pixels thick
A staff line measured three times too thick erased the notes on it and hid 23 whole staves in 185 pages of the 1908 Graduale. The fix and what it changed.
-
A Bank Statement That Misprints Its Own Balance
Six rows of a test statement were still wrong. Glyph widths, whole-number balances, an end marker and a misprint check took it from 56 to 62 of 62 rows.
-
A survey respondent who starts clicking the top button halfway through
Quick pick now keeps the honest start of a respondent who began clicking one screen position partway through, and drops only the run, at under 1% false flags.
-
The tennis feeder that would have tripped its own battery
Sizing the open tennis feeder's fuses found a firmware fault: two launcher wheels starting from rest could ask about 98 A of a 30 A battery.
-
A tilted chant page, read at last, and the bivirga it turned up
On 185 real chant scans, one tilted page went from 2 staves to 9, and checking it exposed a bug that dropped 26 sung notes.
-
A time budget checked between rounds is not a budget
The roster builder's column generation overshot its time share by up to 1.2 s because it read the clock between rounds. The fix was easy; the payoff was not.
-
Four cards on one bill: reading Chase business card downloads
Six public Chase business card downloads: the per-employee Card column we dropped, a spend-by-card report, and a bank label wrong on 17 of 17 lines.
-
The command line reads a receipt photo four times now, and one damaged JPEG taught it something
The ai-bookkeeper CLI reads a small receipt photo four times with tesseract and the readings vote: US photo totals 57.7% to 66.8% right, scans 78.7% to 80.5%.
-
A Rocket Money export booked a firm's card spending as income
Rocket Money writes spending as positive amounts, so four months of a firm's Amex charges imported as income. Monarch's tidy merchant names hid fuel.
-
A scanned bank statement carries its own checksum, so the OCR can be checked
ai-bookkeeper now reads image-only statement PDFs by OCR and fixes misread digits from the running balance: 36 of 40 synthetic scans read perfectly.
-
An elevator interval is not a wait, and the formula's capacity is the knee
The up-peak lift formula matched 20,000 simulated trips within 1.4% in 27 buildings. Waits were half the interval at 80% load, a full one at capacity.
-
Central Park's July nights warmed by 0.79 °C at every quantile; its days did not move
A free earth mover's distance tool compares two groups in your units. On NOAA's Central Park data, July nights moved up at every quantile and days did not.
-
Checking a free loudness meter against EBU Tech 3341, and what Spotify's page actually says
Chant Forge's mix check passed 69 of 69 EBU Tech 3341 readings built from the spec text. Of six streaming services, only Spotify's page gives a level.
-
Adding loudness range to a free mix checker, and the one signal where two meters disagree
Chant Forge's mix check now reports loudness range. It matches EBU Tech 3342's test tones and ffmpeg on 11 of 12 signals; the 12th differs by a percentile rank.
-
Checking a ZATCA stamp without Fatoora
ZATCA's rules check a Saudi e-invoice has a QR code and signature, not that they match it. We now recompute the hash and signature; 45 of 45 samples pass.
-
Colorado Took Its Sales Tax Vendor Fee Back. The 2024 Statute Still Says 4%.
Maryland and Utah timely-filing discounts, read from their statutes, are in ai-bookkeeper's worksheet. Colorado's 4% vendor fee ended in 2026.
-
Croatia's e-invoice rules, and the EU warnings that tell you to delete what Croatia requires
Croatia's B2B e-invoice mandate began in January 2026. Our checker now runs the tax office's HR-CIUS rules and says to keep fields the EU base warns about.
-
Eight sensors, eight chances for noise: the change watch on 33 real pump faults
Pre-registered on SKAB's 33 pump faults: the default watch caught all 33, but on a normal run all 8 sensors together sent 8.5 false alerts per 100 checks.
-
Eight more state sales tax pages read, and not one due date changed
ai-bookkeeper now checks 34 jurisdictions' sales tax due dates on their own pages. Eight read today all matched; NC and VA list no annual filing.
-
Brazil's presumed-credit event has had three schemas in a year
The NF-e event a buyer sends to claim a presumed IBS/CBS credit changed shape twice in a year. Our checker now reads it and converts the old layout.
-
Brazil's reform events: twelve schemas, and one word the manual gets wrong
Every NF-e reform event with a schema is now checked, plus the cancellation. The NT prints one event description differently from its schema.
-
The invoice discount the repair would not touch
The CII repair gave up on any invoice with a header discount or charge. On 66 real ones it now recomputes 64 unchanged and restores 449 of 459 planted faults.
-
Croatia's sample e-invoices fail Croatia's own rules
The Porezna uprava's eight sample eRačuni fail its own HR-CIUS rules, our checker and its XSLT agree 16/16, and our creator now writes HRFISK20Data.
-
We read the laws behind three e-invoicing dates. One was off by two and a half years
Latvia's 2028 B2B date holds, Denmark's opt-out e-invoice enrolment starts 1 March 2027, and the Czech B2G duty began in 2019 and 2020, not 2016.
-
The missing NF-e schema was a space before .xsd
Event 112120's schema is in the official package after all, published with a space in its file name. 211120 is not: NT 2025.002 v1.40 removed it.
-
Four e-invoicing pages rewritten from the tax authorities' own texts: Pakistan, Nigeria, Jordan, Thailand
Pakistan covers every sales tax registrant, Thailand's e-Tax Invoice is a choice, Jordan's regulation sets no date, and Nigeria's portal names UBL, not Peppol.
-
Our Peppol rules were a release behind, and the Danish ones had never run
Moving to Peppol BIS 3.0.22, we ran OpenPeppol's own unit tests on our compiled rules. 124 assertions failed, all Danish, from one compiler bug.
-
Our e-invoice repair changed 69 valid invoices. Now it changes 2.
Run over 529 real, valid UBL, CII and Factur-X files, our e-invoice repair changed 69 and the amount due on 53. Why, and what a repair must leave alone.
-
95 failures from the wrong rulebook: our optional validator checked Saudi invoices against EU rules
Our self-hosted Saxon route ran EU EN 16931 rules on Singapore, Japanese and Saudi e-invoices: 78 false errors on 10 valid files. It now hands them off.
-
Is e-invoicing mandatory in Taiwan? The law says no, and Oman's dates are now settled
Taiwan's VAT Act makes the e-invoice optional, with a 7-day or 2-day upload deadline. Oman's Decision 189/2026 sets 1 April and 1 October 2027.
-
A UAE credit note that must not name its invoice
Under PINT AE, a volume-discount credit note (reason VD) must not reference an invoice. Our checker demanded one. It now reads the reason code.
-
Five chant scan evals, one download queue: a cold run from 303 s to 99 s
chant-forge's scan evals fetched archive.org pages one by one inside their loops. A shared prefetch with six in flight made a cold shape-check 3x faster.
-
Five hundred names for a date column
Our bank CSV reader knows columns by hundreds of names. One pass over the header cells made a header-less file 3.4 times faster, same answers.
-
Four chant endings from the 1908 supplement, and the one that was wrong
Eleven 1908 Graduale transcriptions end off their mode's final. Ten do so because the book does. The eleventh had its last three notes a step low.
-
Four numbers that name a lens: LensSpecification, Nikon's maker note and lensfun
When a raw file names no lens, its focal range and f-numbers at both ends pick one lensfun lens half the time. ShootFlow now uses them for any maker.
-
Freeing eight rounds of a league schedule at once found nothing better, and that is the result
An exact re-deal of up to 8 of 11 rounds could not beat a 12-team round robin's carry-over of 136, and one run reached 136 from scratch.
-
Fujifilm X-Trans raw files, without the converter
ShootFlow now develops Fujifilm RAF files itself, X-Trans included. Every sensel matches LibRaw on three bodies; the colour layout is stored backwards.
-
Indiana's sales tax rules look at your last year, so the worksheet reads it from your books
Indiana's collection allowance and 20th-or-30th due day depend on your past tax, so the free bookkeeper now works both out from your books.
-
Screen printing a painting: the inks for dots are not the inks for flat colour
Average colours make poor inks for an index print. Inks chosen for the dots, grain counted, cut the error at 3 m by a quarter on nine paintings.
-
Is 136 a wall? A faster refill search finds 210 at 14 teams and nothing new at 12
A refill search in C found a 14-team round robin at carry-over 210, below 212, while 631 fresh climbs at 12 teams all stopped at 136 or higher.
-
January 31, 2026 is a Saturday
Filing and paying by January 31 forgives the January estimate's penalty. That day moves past a weekend, so for 2025 returns it is February 2, 2026.
-
Lensfun's 16-bit vignetting draws stripes every four pixels
We ran lensfun 0.3.4's float and 16-bit vignetting code through ctypes. Its 16-bit SSE2 loop draws 4-pixel stripes and breaks each row's end.
-
Mustang passed our Factur-X PDFs, then pointed at one empty element
Mustang 2.26 called all eight Factur-X PDFs from our free writer valid, but warned on each one. Both warnings came down to a missing delivery date.
-
$9,000 collected, $2,890 earned: deferred revenue for small subscription businesses
AI Bookkeeper now books annual plans and retainers as earned, by day or month, with cancellations, refunds and a current/long-term waterfall.
-
The dot a screen can't print: taking single dots out of an index separation
Error diffusion leaves 6-9% of cells as lone dots a mesh may drop. Removing them, then searching under that rule, kept prints closer in 33 of 36 runs.