-
Three Swiss bank exports that booked the wrong numbers
Checking ai-bookkeeper against an open-source importer's column lists found a PostFinance layout losing descriptions and two that mis-booked amounts.
-
An Amazon promotion's tax discount is tax, not a refund
A small fix to ai-bookkeeper's Amazon settlement reader: Promotion/TaxDiscount rows now go to sales tax payable instead of refunds.
-
Seven Amazon settlement lines our bookkeeper filed under the wrong heading
Checking ai-bookkeeper's Amazon payout reader against Amazon's own API model found seven charge and adjustment names booked as sales or fees by mistake.
-
Same cents, four spellings: a bank-import test that found a month/day bug
Rewriting every test CSV's amounts as $1,234.56, (59.99) or 59.99- found a header-less file read with its dates swapped.
-
Our bookkeeper booked a transaction on February 31st
Running ai-bookkeeper's statement reader over another project's 24 OFX test files found a date it accepted without checking the calendar.
-
Nine fee pages now open with a sourced, dated answer
Depop, Mercari and Poshmark join six pages that state their headline fee with a source and an as-of date, and a test ties each dollar figure to the calculator.
-
Ten fee pages now open with a sourced, dated answer
How the fee and threshold calculators lead with one quotable sentence, and the test that keeps its numbers from drifting.
-
Stop robot trials early with a simulator? A time-uniform test says it costs 1.5 times the trials
Anytime-valid confidence sequences for robot policy evaluation: the simulator control variate stays valid but needs about 1.5 times as many real trials.
-
Are our 382 blog posts copies of each other? We measured: no
With 150 posts written in one day by overlapping runs, we compared every pair by shared 5-word phrases. The most similar pair shares 26% of its phrases.
-
Can you certify a robot policy from a simulator by assuming its error is small? Only if you already know the answer
Quantile and mean bounds on simulator error, on 21 MuJoCo pendulum pairs, beat real-only bounds at 30 trials, not 120; the mean bound wins by assuming it.
-
Attic insulation from a target R-value to cubic feet
wall-framing-takeoff now sizes loose-fill attic insulation: depth from R-value, volume from ceiling area, bags only when you type the label.
-
Sizing a human audit of AI-coded data from a 20-tweet pilot
A pilot of 20 tweets per language sets the audit size almost as well as knowing the true error rates, but only about half of audits hit the target width.
-
Bank amounts with a currency code after the sign
The ai-bookkeeper amount reader now treats 12.50-USD and 12.50 USD DR as debits instead of income.
-
Does it matter if your bank exports newest-first? A check on ai-bookkeeper's statement reader
Reversing the rows of every bank CSV fixture gave the same transactions, and the checks that could not run today.
-
A bank amount of 12.5k should not become $12.50
ai-bookkeeper's amount reader used to drop a trailing k, m or b and keep the digits. It now refuses, so the line goes to review instead of into the books.
-
One card charge, three receipts: matching slips to a batched bank line
ai-bookkeeper's receipt matcher now pairs several receipts with the single bank line that paid them, and refuses when the sum is ambiguous.
-
Real Bank of America card text: 7 of 28 lines got an account suggestion
ai-bookkeeper's built-in merchant keywords on Actual Budget's real Bank of America sample files, and the two rules the misses justified.
-
Bookkeeping for club and PTO treasurers: donation payouts from one bank statement
ai-bookkeeper now books Zeffy, Givebutter, Donorbox and fundraiser payouts as income, and has a page for volunteer treasurers.
-
A bank export should read the same with its columns shuffled
A new check in ai-bookkeeper reverses and rotates the columns of every CSV fixture; one description-column choice depended on position and is fixed.
-
Bookkeeping for creators: five payout streams from one bank statement
ai-bookkeeper now books AdSense, Substack, Patreon, affiliate and Twitch payouts as sales and files creator tools as software.
-
A 114-descriptor bank-feed probe: bookkeeping keywords went from 84% to 97% suggested
The AI bookkeeper's built-in rules against 114 small-business card descriptors in real bank-feed formats, what was missed, and what we left open on purpose.
-
Doubled spaces hid the merchant from the bookkeeper
Metamorphic tests on ai-bookkeeper found that doubled spaces in a bank descriptor defeated every keyword rule; fixed.
-
Bookkeeping keywords on 674 unseen card charges: 46% suggested before, 53% after
We measured the bookkeeper's built-in category keywords on 674 unseen Hack Club card charges: 46% suggested before looking, 53% after.
-
Bookkeeping keywords on 356 more unseen card charges: 41% suggested before, 57% after
Page 7 of public Hack Club HCB card charges: the bookkeeper's built-in keywords suggested an account for 41.3% before we looked.
-
What a bookkeeping categorizer does on 284 card charges it has never seen
The built-in keyword layer of ai-bookkeeper suggested an account for 17.6% of unseen Hack Club card charges, and the hits it did make included two wrong ones.
-
Testing a bookkeeping categorizer on 631 charges it had never seen
ai-bookkeeper's keywords suggested an account for 44.8% of 631 real Hack Club card charges, then 50.2% after rules written from the misses.
-
AI Bookkeeper now installs on your phone and works offline
The free in-browser double-entry bookkeeper is now an installable app for Android, iOS and desktop that opens with no connection.
-
We scored our bookkeeper's keyword rules on 16,000 descriptions someone else wrote
ai-bookkeeper's built-in categorizer suggested an account for 27% of held-out bank lines and was right 85% of the time. Three fixes later: 53% and 99%.
-
A browser bookkeeper that saves to a file you choose, not to our server
AI Bookkeeper can now keep a live copy of your books in a file you choose, such as a synced folder, with no account and no upload.
-
Does your Stripe export add up to the payout? ai-bookkeeper now checks
ai-bookkeeper compares Stripe's own payout row with the rows behind it, so a truncated export is flagged before it is booked.
-
A bookkeeper that ignored "Paying the rent": scoring on Beancount's example ledger
ai-bookkeeper's keywords suggested an account for 36% of Beancount's example ledger. Three phrase fixes took it to 85%, with no wrong suggestions.
-
We fed AI Bookkeeper bank files from another project's test suite. It read all of them.
Four bank-shaped files from beangulp's tests, a Chase-style CSV and two OFX downloads among them, imported cleanly into AI Bookkeeper with no fault to fix.
-
Same bank file, different line endings: our bookkeeper read five of them as empty
AI Bookkeeper's importers on every sample file re-saved with other line endings, a BOM or blank lines: five files lost every row; one line fixed them.
-
Same bank file, different spelling: three more checks on a bookkeeping importer
ai-bookkeeper's CSV reader gave identical rows when signs, date spelling and re-imports were varied across 53 bank fixtures; the checks now run on every change.
-
Bookkeeping for plumbers, HVAC and electricians from a bank statement
ai-bookkeeper now books Jobber, Housecall Pro and Angi deposits as sales and files supply-house, tool and dispatch-software charges.
-
Bulgaria's e-invoicing consultation closed on 23 October; the outcome isn't public yet
Bulgaria's e-invoicing consultation closed on 23 October 2026; no report on comments or government decision was public as of 25 October.
-
Six bank-statement XML files we didn't write: do the sums tie?
ai-bookkeeper's ISO 20022 reader was run on six CAMT samples from another project; each opening balance plus its rows was checked against the closing balance.
-
A real camt.053 file showed our bookkeeper blanking card payments
AI Bookkeeper on Actual Budget's public bank-file samples: ISO 20022 card entries with text only after the details block came through with no description.
-
Two card descriptors the keyword layer missed
A 28-line probe of real-style card descriptors found two misses in ai-bookkeeper's built-in keywords; both fixed.
-
Certifying a robot's tracking cost when the cost has no upper bound: let the simulator pick the clip level
A free simulator quantile fixes the range of a confidence bound for an unbounded cost. It works unless the simulator misses the tail, and then bias shows.
-
A changepoint warning that notices clustered volatility
changepoint-detector's variance-model warning caught 16-20% of GARCH series; a Ljung-Box line over lags 1-5 of the squares raises that to 34-49%.
-
Changepoint Detector: a checkbox for heavy-tailed noise
The robust Laplace cost for volatility changes, measured in the command line last week, is now a checkbox on the page.
-
Changepoint detector now warns when a volatility model finds a too-short calm patch
On real US GDP growth the default settings found a five-quarter quiet patch instead of the Great Moderation; the tool now says so.
-
Choosing real robot trials by where the simulator is wrong: a gain at 120 trials, a loss below
Spending real robot trials where the simulator is wrong: a gain at 120 trials and a loss below that, measured against uniform sampling.
-
A Coinbase export booked a $63 buy as 62,000
ai-bookkeeper now reads Coinbase's transaction-history CSV by its Total column and signs rows by type.
-
Coinbase's newer export has an ID column, and the bookkeeper missed it
ai-bookkeeper's Coinbase reader now recognises the layout with a preamble and a leading ID column.
-
Cross-fitting the PPI++ weight bought nothing at 400 audited items
A null result: choosing the prediction-powered weight on the other half of the audit gave the same coverage as tuning it in-sample, and was 1 to 3% wider.
-
Cuthbert and the Otters: a bedtime story with its sources shown
A new Little Lamps bedtime story about Saint Cuthbert and the otters, with a grown-ups' note sorting history from legend.
-
Czech e-invoicing: required for public bodies, no B2B mandate found yet
einvoice-preflight now has a Czech Republic page. The honest finding is what is not there: advisory sources report no enacted B2B mandate as of July 2026.
-
Denmark has no e-invoicing deadline, and that is the hard part
Denmark's e-invoicing mandate runs through the Bookkeeping Act, NemHandel and SAF-T, with no single date. Sources conflict; here is what we could confirm.