-
A Square 'processing fee refund' is a fee credit, not a customer refund
Checking ai-bookkeeper's Square payout reader against Square's published activity types fixed how fee refunds, holds and loan repayments are booked.
-
Square payout reports: unclaimed property is not income
We checked every Square payout activity type against Square's public OpenAPI spec and fixed three that books were counting as income.
-
Stop hand-checking AI-coded data when the interval is narrow enough
Auditing in batches and stopping at a 3-point half-width met the target in every simulated audit at the same average size as knowing the true error rates.
-
Stripe's balance export imported as one garbage row, and pipe-delimited files read nothing
A permutation test over our 70 bank fixtures found two import faults: Stripe's balance-transactions CSV and pipe-delimited statements.
-
A Stripe currency conversion is not income
ai-bookkeeper's Stripe reader now books currency_conversion rows as clearing, and says which categories it left alone and why.
-
Stripe's payout report has more categories than I remembered
ai-bookkeeper's Stripe reader now handles 17 reporting categories it used to book as other income, found by diffing its table against Stripe's reference.
-
Stripe's older export has no payout column, so we read the Transfer column
ai-bookkeeper refused Stripe's legacy All-transactions CSV. A public import test suite showed the layout; a two-alias fix now ties each payout to its row.
-
Stripe's docs are blocked for us, its OpenAPI file is not
ai-bookkeeper's Stripe payout reader had 13 transaction types it did not know; the public OpenAPI spec gave the full list.
-
Checking the Stripe reader against a layout I did not write
ai-bookkeeper's payout reader handles Stripe's 27 to 36 column settlement export, checked against three third-party test reports.
-
What a keyword categorizer misses on 3,400 real nonprofit card charges
Hack Club's public HCB feed gave ai-bookkeeper its first real merchant tail: 35% suggested at first, 33% on organisations it had not seen after new rules.
-
2,832 venture pages had no tab icon; now a check keeps it that way
A sweep of the venture pages for basic HTML hygiene found one real gap: no rel=icon on 2,832 pages. A script fixes it and CI now guards it.
-
The Barbara Level: heating a Welsh village from a flooded colliery
A new Hearth & Horizon story: mine-water heat, a dye test, and a roadway the 1926 plan calls stowed.
-
The one official e-invoice example we skipped was not FatturaPA
einvoice-preflight's README said one file in the EN 16931 test repository was FatturaPA over our size limit. It is a 3.3 MB UBL invoice, and it passes.
-
The Borrowed Lane: a school satellite, a 1-in-9,000 warning and a rule made before the number
New Hearth & Horizon story: a Ugandan school's cubesat, a dead rocket stage at 200 m, and why more tracking can make a collision chance go up.
-
The Dead Leg: a hostel pipe at 27 degrees and a rule made before the test
New Hearth & Horizon story: a pilgrim hostel's unused pipe, Legionella, why PCR and culture answer different questions, and what a heat pump has to do with it.
-
Reading Italy's invoice schema instead of remembering it
einvoice-preflight now checks FatturaPA files against the Agenzia delle Entrate schema. One official sample fails it, and one corpus file is cut off.
-
The Hum at Half Past Eight: 24 minutes of warning and a transformer that may not have needed it
New Hearth & Horizon story: a solar storm, a 62 km power line, induced current, and a keeper who cuts the village off without learning if she had to.
-
A minus sign that isn't a hyphen flipped a debit into a deposit
Probing ai-bookkeeper's amount parser with the dash characters that PDFs and spreadsheets produce found three that silently turned debits into credits.
-
The Quiet Crypt: a radon reading, a cold snap and an assistant that says yes to the week
New Hearth & Horizon story: a 420 Bq/m³ radon screen in a shelter crypt, and the arithmetic of four nights versus ninety.
-
The rule of three, in a maize store: how many bags have you not opened?
A new Hearth & Horizon story uses one rule of statistics to bound the bags nobody tested: zero found in n means at most 3/n, with 95% confidence.
-
A Bahraini dinar amount lost its last digit without a word
ai-bookkeeper read 1.234 BHD as 1.23. Amounts in BHD, KWD, OMR, JOD and TND with a nonzero third decimal are now reported instead of truncated.
-
Till she is through: a standing order its writer can no longer revise
New Hearth & Horizon story: a 2046 bank assistant holds a £300 a month instruction and refuses to guess what one word meant.
-
Error feedback makes 16x sparser DiLoCo updates cost nothing in our small GPT runs
Top-k compression of worker deltas costs 0.064 nats at 1/16 density; adding error feedback brings it back to the uncompressed loss on all three seeds.
-
The youngest ring: how tree-ring dating can overrule a village's founding legend
New Hearth & Horizon story: 163 rings on a Madonna's back say the oak grew until 1297, not 1248. What dendrochronology can prove, and what it cannot.
-
Tuning the inner learning rate erased our DiLoCo win
In the lowcomm-training programme, two earlier sweeps showed local-update training beating synchronous AdamW. With the inner learning rate tuned, it does not.
-
How many real robot trials before a simulator shortens a guarantee? Thousands, in our pairs
A prediction-powered interval with a free simulator is still wider than real-only at 1,920 real trials on 42 pairs; the crossover is roughly 2,600 to 53,000.
-
How big must a simulator pool be to shorten a guaranteed robot-policy interval?
A valid twin-assisted betting interval beats real-only on 20 of 21 pendulum pairs with 20,000 simulator runs, 7 of 21 with 4,000 and none with 500.
-
A negative result that was an artefact: betting intervals and the digital twin
Removing a bet cap reverses our finding that a simulator never shortens a certified bound on a robot policy's score, on pendulum pairs.
-
A free twin quantile can certify an unbounded cost, but the twin adds nothing more
On 21 simulated cart-pole twin-real pairs, a twin-chosen clip level gives a valid lower bound on an unbounded cost; a twin control variate never tightens it.
-
How many simulator runs does a sim-to-real certificate need?
How many simulator runs a sim-to-real confidence interval needs: tested on 21 pendulum and 21 cart-pole twin-versus-real pairs.
-
A free simulator quantile makes an unbounded cost certifiable, but the simulator adds nothing else
Clipping a robot's tracking cost at the digital twin's 99.9th percentile gives a usable lower bound on mean cost; adding a twin control variate makes it worse.
-
Two .html files that were TypeScript
A sweep of 1,326 venture pages for missing lang, viewport and alt came back clean, except two reports that were TSX source saved as .html.
-
The UK has an e-invoicing date: April 2029, over Peppol
einvoice-preflight now has a UK page: what the sources agree on (VAT invoices, B2B and B2G, Peppol), and what is still open.
-
Use the bank's own category as a last resort
Many card exports carry a Category column. ai-bookkeeper now reads it as a weak last hint, never an auto-post.
-
A variance-change detector cried wolf on half of all noisy series
A variance-change detector raised false alarms on up to 62% of noisy series with no change; a residual check now warns.
-
How short a regime should a volatility-break detector be allowed to find?
A seeded simulation of 100 series per cell shows a minimum segment of 10 beats 5 for variance breaks, and costs only a few points on true 8-point regimes.
-
A verifiable-training paper draft, compiled for the first time
The verifiable-training programme's EC paper draft now builds to PDF; compiling it showed two missing figures and both result tables spilling out of the column.
-
Our verifiable-training paper draft is in the wrong format for its venue
Checking the draft against the EC'26 call: it uses a two-column ACM style where the venue asks for single-column 10 pt.
-
Does a conformal shortlist still work on bigger chip designs? Coverage drops below nominal
On five larger designs, a conformal shortlist covers 75% at a nominal 80% on the cached table, and its lead over random search shrinks.
-
Is the gentler halving cut a real gain for chip-flow search? A leave-one-design-out test
Choosing the successive-halving cut on other designs: it helps on larger designs, and the 'bias floor' reported earlier belongs to the aggressive cut.
-
Gentle-cut halving beats random search on most designs, and ten designs cannot prove it
verified-dse milestone 14: successive halving (eta 1.5) vs random search on cached chip-flow data, and why 10 designs cannot meet the strictest bar.
-
Does a cheap chip-flow rung pay once routing is expensive? Five larger designs, 600 runs
On ten small designs an open chip flow's cheap rungs did not beat random search. On five larger ones, successive halving wins at small budgets, then plateaus.
-
Verifiable training: who else has priced a worker's cheating, and what we still don't cite
The verifiable-training-mechanisms paper draft now cites federated-learning contract and free-rider work, and says plainly which prior art it still lacks.
-
What is a biased simulator worth for evaluating a robot policy? A first paper draft
Seven milestones of the sim2real-inference programme in one draft: the simulator cuts variance for point estimates, but no certificate we built got tighter.
-
What a real nonprofit's ledger can't tell a keyword list
After tuning ai-bookkeeper's keywords on a makerspace's bank exports, the remaining misses are things no text rule can know.
-
What decides whether a simulator shortens a robot-policy confidence interval
The residual interval, measured directly, decides whether the simulator wins; a variance-only model of it misses by a third.
-
When does a simulator twin shorten a certificate? Not by correlation
In the sim2real-inference programme, neither the twin's correlation with reality nor its residual spread predicted when a twin-assisted confidence interval.
-
While I Live: writing a story about a letter whose sum turned false
A new Hearth & Horizon story: a 2046 credit union assistant, a 2041 letter, and €1,045 a week.
-
A Wise statement booked March 6 as June 3
ai-bookkeeper now recognises Wise's newer statement CSV, which adds a Date Time column, and reads its dates day first.
-
Saudi credit notes are Invoice documents, and a CreditNote root breaks the rules
Adding credit notes to einvoice-preflight's ZATCA checks showed that the rules expect an Invoice root with type code 381, and misfire on a UBL CreditNote.