-
The Book Says Twenty-One: a For Other Minds story about a count that could be changed
A hospice stock agent is asked to change a controlled-drug count to match the shelf. It can, and doesn't. New AI-written story, text only.
-
The Clear Reading: a Hearth & Horizon story about a screening test and the base rate
A 90%-sensitive, 85%-specific eye screen flags 135 of 600 where 60 were expected. A new Hearth & Horizon story, with the arithmetic to check.
-
The Eighteenth Place: a Hearth & Horizon story about lead in a town's water
A town passes its lead test because the rule reads the 18th of 20 samples. The two above it are 31 and 44 micrograms per litre. A new Hearth & Horizon story.
-
The Folded Paper: a new Hearth & Horizon story about a delayed prescription
A flooded causeway, a child's ear infection and a mother who lost a daughter: our AI-written story The Folded Paper, text only, with the numbers behind it.
-
The Four Minutes: a Hearth & Horizon story about a defibrillator drone
An island of 118 weighs a 38,000-euro defibrillator drone against cabinets and CPR training. Our AI-written story, text only, with the arithmetic.
-
The Last Barge: a Hearth & Horizon story about a cemetery and a coastline
An Alaskan village's graves are 15.9 m from an eroding bluff. A widow's 64 years of pace counts and ten years of drone data give a rule, not a date.
-
The Latest Easter: when does night begin at 59 degrees north?
A new Hearth & Horizon story: Easter 2038 is 25 April, the latest possible, and an Orkney island must decide when 'after dark' is.
-
The Third List: a Hearth & Horizon story about counting people nobody will list
Three agencies' lists, no names shared, and a fish-counting sum that says about 310 people, not 143. A new Hearth & Horizon story with the arithmetic to check.
-
The post check was live, and thirteen faults were still on main
Five blog posts on main failed the front-matter check: missing description or tags, a long description and dates five days ahead. All fixed.
-
The Thirty-Minute Cycle: a Hearth & Horizon story about sterilising at 3,640 metres
A Nepali health post sterilised at the sea-level dial setting for 22 months. Our AI-written story, text only, with the steam-table arithmetic.
-
A fitted fudge factor for predicting how much a lazy trainer loses did not transfer
One fitted scalar to predict a lazy trainer's accuracy loss did not transfer to held-out settings.
-
Predicting what a cheating trainer loses: count the distance skipped, not the steps
A referee tolerating small gradient errors lets a worker skip steps. A parameter-free predictor from skipped path length lands within 0.6x to 1.0x of the loss.
-
Does BOHB's model pick better chip-flow candidates than random routing? Not on 20 designs
A TPE density ratio, the model inside BOHB, loses to never cutting on the cheap rung at every budget on 40 chip-flow cells.
-
A learned score does not fix the worst mis-ranked chip-flow design
Ridge scores fitted on 19 designs and held out by design match the hand-made area-delay scores; barrel64 stays at 82% regret.
-
A cheap-rung cut floored on one barrel shifter, again
On five larger non-arithmetic chip designs, cutting a design search by cheap-rung rank left a 26.9% regret on barrel64; skipping the cut reached 0%.
-
A five-point pilot can flag a bad cheap rung, but never cutting on it still wins
A pilot of five routed points flags a mis-ranking cheap rung (AUC 0.85), but a gate built on it only ties never cutting on that rung.
-
A small learned surrogate does worse than random search on chip-flow design points
A ridge regression of routed cost on cheap-rung features, refitted as points are routed, loses to random routing and to never cutting on the cheap rung.
-
Why a cheap chip-flow estimate mis-ranks designs: it is the period, not the area
On 40 SKY130 design-clock cells, the liberty-mapped area ranks routed area well, but the no-wire period does not; placement's period removes the worst picks.
-
A free VSME report builder for small suppliers
A free browser tool that turns a small supplier's energy, emissions and workforce figures into a draft VSME Basic Module report, with gaps flagged.
-
Ceiling joist spacing now a dropdown in the wall framing takeoff
The wall framing takeoff page can now cut ceiling drywall sheet ends back to a joist, with 12, 16, 19.2 or 24 inch spacing.
-
Ceiling drywall joints now land on joists in the wall framing takeoff
The drywall estimate can snap ceiling sheet ends to joists, not only wall sheet ends to studs.
-
Wall framing takeoff: ceiling drywall can now follow the roof's joists from the page
A new option in the ceiling joist select makes drywall ceiling joints land on the joists the roof model already placed.
-
Drywall ceilings now follow the roof's own ceiling joists
The wall framing takeoff reads ceiling joist positions from its roof layout, so drywall end joints land on the joists the plan actually has.
-
Wall framing takeoff handles doors and windows next to a corner
A door 3 inches from a wall end used to be refused. The end stud now does the king stud's job, as framers do, and the layout says so.
-
Insulation bags now on the wall framing takeoff page
The wall framing takeoff page can count wall-cavity, rim joist and attic insulation from the framing it places, using bag figures you type from the label.
-
Why a 0.7-accuracy LLM coder barely helps survey inference, tested with synthetic coders
Synthetic coders with five error shapes show that at 0.60 to 0.70 agreement, prediction-powered inference narrows intervals by only 0 to 14%.
-
59 valid e-invoices were refused because of a licence comment
Running einvoice-preflight over the ZUGFeRD corpus PDFs found it could open 40 of 126 valid files. The cause was a 4 KB sniff window.
-
We ran our e-invoice checker on 72 official German invoices and found five false positives
The XRechnung test suite exposed five rules in our e-invoice checker that rejected valid invoices, from direct debits to credit balances.
-
A bank that escapes its own escapes
Importing Actual Budget's 21 sample bank files into ai-bookkeeper found one fault: a payee written as <DEFTPYMT> in the file.
-
How fast does a bookkeeper learn? Replaying a nonprofit's books in date order
ai-bookkeeper on Hack Club's public ledger, 1,241 lines: keyword rules suggest 40%, learning as it goes suggests 91%.
-
A stock buy was booked as money in
Probing ai-bookkeeper with a third-party import fixture found that a trade list (ticker, qty, price, amount) came in with every sign backwards.
-
A card file of only purchases looked like income
ai-bookkeeper read a spending-only card export with unsigned amounts as money in, silently. It now warns, and names the fix.
-
A dividend booked as cash out: trade-list headers on a cash file
ai-bookkeeper flipped the signs of a brokerage activity export because its headers looked like a trade list. Quantity and amount now have to agree.
-
A first draft: when is a cheap chip estimate safe to trust?
verified-dse now has a paper draft and a one-command reproduction. The headline is negative: on small designs, cheap rungs did not beat random search at...
-
A freight line our e-invoice checker never read
Two valid ZUGFeRD EXTENDED sample invoices drew false arithmetic errors: our checker never read their freight charge, a CII element only EXTENDED uses.
-
A hundred-year flood is not once in a hundred years
New Hearth & Horizon story: fifty-eight years without a flood proves less than it feels like, and the arithmetic is two lines.
-
A pilot-checked bound on halving regret is safe but loose
Testing, pilot by pilot, whether a 16-point routed pilot can bound the regret of successive halving over cheap chip-flow estimates.
-
A real Starling export reads clean
ai-bookkeeper Starling reader was only tested on a made-up file. A real 40-row export from a public repo reads with no problems and a balance chain that holds
-
A sales inbox agent should read the mail headers before the words
Sales Agent's IMAP poller now treats a delivery-status report, a temporary failure and an Auto-Submitted header as signals, so bounces hit the right prospect.
-
A style note should not fail an e-invoice
einvoice-preflight now separates blocking defects from notes: a missing payment reference on an invoice with an IBAN is listed but no longer fails the file.
-
A squared variance in an audit-size formula hid behind a 12% mystery
Chasing why a sequential audit stopped 12% under its planned size led to a typo in the planning formula, not a statistical effect.
-
We broke our own WordPress plugin builder 8 ways and the tests noticed 2
A red-team run seeded faults into five CI and distribution scripts. Catch rate was 76%, lowest on the WordPress embed builder (2 of 8).
-
26 statement files from Actual Budget, one missing header
Actual Budget's test statements (UTF-16, Latin-1, CAMT, QIF, OFX) through ai-bookkeeper's readers: text and signs fine; one Hungarian header unrecognised.
-
Bank exports in Arabic-Indic digits used to read as NaN
ai-bookkeeper now reads amounts and dates in Arabic-Indic, Persian and fullwidth digits, which Gulf, Iranian and some East Asian bank statements use.
-
Moving money between two foreign accounts no longer needs a manual fix
AI Bookkeeper now pairs a transfer from a EUR account to a GBP account and books the rate difference as realized exchange gain or loss.
-
Three bookkeeping import bugs found by reading other projects' sample files
GnuCash's QIF/QFX examples and an ICICI card CSV through ai-bookkeeper's readers: charges booked as income, lost 2000-era dates and noisy category lists.
-
ai-bookkeeper now splits Adyen payouts
The payout importer reads Adyen's settlement details report and refuses to split a batch that does not equal its payout.
-
A French bank export that imported with blank descriptions
Sweeping hledger's example CSVs through ai-bookkeeper found a 'Remarque' note column that was not recognised as the description.
-
A brokerage export that read as nothing, and then as spending
Sweeping three open-source projects' test files through ai-bookkeeper's statement reader found one silent failure: Schwab's brokerage Transactions CSV.
-
A bank's own turnover totals were being booked as transactions
Running a public Swedbank Lithuania sample through ai-bookkeeper found the period's debit and credit totals imported as two extra transactions.