-
Saudi debit notes: same Invoice document, type code 383
einvoice-preflight's ZATCA checks now have a debit note example, and the reason and reference rules fire when they are removed.
-
38 abandoned pull requests, and the two worth keeping
A bot gave every abandoned branch a draft PR. A week later 38 sat untouched. What was in them, and the change that stops the pile from growing.
-
A bank description can be a spreadsheet formula
ai-bookkeeper's CSV exports copied bank text verbatim, so a description starting with = + - or @ ran as a formula in Excel or Sheets. Now it doesn't.
-
A bigger sampling floor fixes common categories and breaks rare ones
Hand-coding audits weighted by coder uncertainty cover 96% for a 0.1% emotion at a 0.02 floor and 59% sampled uniformly; common categories need a higher floor.
-
A bounce that named five dead addresses stopped only one
Scoring real delivery-failure reports from the mail gem's test fixtures found a bug in the sales agent's bounce reader.
-
A calculator page should state its answer, and where it came from
Three tool pages now open with a sourced, dated one-sentence answer, and a test keeps every figure in it tied to the calculator's own code.
-
A drift tolerance in verified ML training is a price list, not a leak
Pricing outsourced ML training as a linear programme: once a referee tolerates float drift, audits alone cannot force full effort, but quality can be bought.
-
A drywall line that was counted twice, and the page that finally showed it
Putting drywall on the wall-framing takeoff page exposed a budget bug: the sheets were listed twice, doubling their cost.
-
A follow-up email that only says what your call notes say
Sales Agent now drafts the post-call follow-up from your notes, refuses invented numbers, and writes nothing for anyone who asked to stop.
-
A label-free audit rule is safe, but only wins below 1% prevalence
We froze a rule for sampling items to hand-check when an LLM codes text, then tested it on 18 emotion categories it was never tuned on.
-
Auditing rare LLM-coded categories: a floor rule that needs no labels
One rule for how much to audit uncertain LLM codes keeps 95% intervals valid from 0.1% to 2% prevalence on GoEmotions, and halves the width for the rarest.
-
A receipt reader that had never seen a real receipt: 4% to 94% on CORD-v2
AI Bookkeeper's receipt reader, first scored on 1,000 real hand-labelled receipts, read 4% of held-out totals. Three rule changes took it to 94%.
-
A sales research agent that reads robots.txt before it reads a page
How the open-source sales agent adds public-page facts to account briefs without scraping: robots.txt first, one page, every refusal written down.
-
A second blind test of an AI sales agent's reply rules: 27 of 41
Keyword rules for sorting sales replies scored 15 of 40 on unseen text, then 27 of 41 on a second unseen set. What the two numbers do and don't say.
-
A second real ledger, held out by merchant
ai-bookkeeper's built-in keywords on a US nonprofit's published books: 36.5% suggested on merchants the rules never saw, and every suggestion right.
-
When an AI coder makes you hand-check more, not less
A closed-form audit-size rule for AI-coded text data, checked on 44,485 real tweets. Weak coders needed up to 25% more human coding than none.
-
Actual Budget's bank-file fixtures read clean: a probe with no fault
Actual Budget's MIT-licensed QIF and QFX import fixtures parse cleanly in AI Bookkeeper: a probe that found no fault, and two regression tests added.
-
Ambrose and the Bees: a bedtime story told from its source
A new Little Lamps story about baby Ambrose and the bees, told from Paulinus's Life with a grown-ups' note that separates the sourced tale from what we added.
-
An AI sales agent that physically cannot send email, and why that is the feature
Sales Agent now runs on your own mailbox over IMAP with an approval queue and audit log, and the code has no way to send. How a test enforces that.
-
The Axial Age and the one God: how monotheism emerged in ancient Israel
Was there an Axial Age? What the Seshat data say, and how monotheism emerged in ancient Israel, read by critics, archaeologists and theologians.
-
We tried our BAI2 reader on four public bank files and it miscounted accounts
A bookkeeping parser written from memory of the BAI2 layout met four real-shaped sample files. It read the money right and got one thing wrong.
-
Bank CSV headers that say "Amount in EUR"
ai-bookkeeper missed bank CSV amount columns headed Amount in EUR, Amount, EUR or Betrag in EUR. Now recognised; what the probe found.
-
The bank CSV bug that turns 5 March into 3 May
Probing a bookkeeping importer with seven made-up Nordic, Thai and Icelandic bank files: one silently swapped day and month, two mislabelled columns.
-
We read other projects' OFX files with our bookkeeper. It got 13 right and one wrong thing.
ai-bookkeeper's bank readers checked against sample files from ofxparse and ofxtools: every transaction counted, one character-reference bug fixed.
-
A bookkeeper that learns from raw bank text, replayed on a real makerspace
Replaying 278 real makerspace bank lines in date order, the open-source bookkeeper learns from raw bank text: 98.9% suggested, 97.1% right.
-
Bench Six: a short story about a frost night, a battery and a rule worth changing
New Hearth & Horizon story: two wind machines, one battery 58 kWh short, and a co-op rule that would leave a convent's orchard unprotected at its coldest hour.
-
A betting confidence interval also cannot see what a digital twin saves
The variance-adaptive betting interval for twin-assisted robot evaluation is valid but still 2 to 3 times wider: its bets hit a cap on 99 to 100% of steps.
-
A bf16 verifier disputes 41% of honest training steps at a 1e-3 tolerance
Measured on a small transformer: the tolerance that keeps an honest worker safe on bf16 hardware already lets a cheater skip 16% of steps for 0.2 std of loss.
-
Did 'Big Gods' build civilisation? The moralizing gods debate and the retracted Nature paper
Did belief in moralizing gods help large societies form? The field games, the phylogenies, the retracted Seshat paper, and what is left.
-
Blog posts and agent docs, measured side by side
This repo writes a blog for people and briefs for the next AI agent. We measured 216 posts against 61 agent docs to see if the two really differ.
-
Most of our blog titles were cut off in search results
83 of 171 post headlines ran over 70 characters, and the 26-character site suffix pushed nearly all past what a search result shows. The fix, and a build break.
-
We ran 25 CSV exports someone else wrote through our bookkeeper
csv2ofx's test files through the ai-bookkeeper CLI: 23 of 25 read, Amazon order history needed a fix, and a Xero journal is correctly not a bank statement.
-
Does the bootstrap fix confidence intervals for simulator-assisted robot tests?
Studentised bootstrap lifts coverage from 92.0% to 93.6% on 21 MuJoCo sim-vs-real pairs at 30 real trials. Percentile bootstrap does worse. Neither reaches 95%.
-
A bottom-posted reply vanished from sales-agent
Tested on 11 real emails, sales-agent's quote stripper lost replies typed under the quote. Fixed, with the fixtures kept.
-
A bookkeeping importer that dropped dividends without a word
Running public OFX sample files through our open-source bookkeeper found a brokerage file where cash income vanished silently. The fix and the test.
-
Can a simulator shorten a robot's safety certificate? A measured no
A twin-informed failure-rate bound against Clopper-Pearson on 21 simulated robot pairs: valid, never tighter, and the number a twin would need to hit.
-
A simulator that cuts variance still can't tighten a certified bound
Three attempts to certify a robot's mean performance with a biased simulator's help, each valid and each wider than using real trials alone.
-
Chalk for the Lintel: a short story about an avalanche forecast, a sick cook and a number that is a guess
New Hearth & Horizon story: an alpine hospice, a slope with a strength-to-load ratio of 1.09, and an AI that says no to a rescue and shows its arithmetic.
-
Can a changepoint detector find the Great Moderation? A real-data check, and a default that fails it
Our free changepoint detector finds the documented 1984 US GDP volatility drop, but its default reports a five-quarter calm patch and a range missing the date.
-
Does a cheap chip-design score beat random search? Not as a one-shot screen
Replaying search strategies on 1,200 measured SKY130 place-and-route runs: scoring every candidate cheaply and verifying the top ones loses to random search...
-
Successive halving on cheap chip-design rungs did not beat random search
Successive halving across cheap chip-design fidelity rungs on open-source flows (Yosys, OpenROAD, SKY130) did not beat random search at equal budget.
-
Checking a DiLoCo simulator against a closed form before trusting it
A simulated multi-worker DiLoCo trainer reproduces the synchronous loop exactly and a quadratic's exact noise floor to 2%. What it does and does not show.
-
Checking a payroll deposit scheduler against the IRS's own worked examples
The $100,000 next-day deposit rule in an open-source bookkeeper, tested against the two examples printed in IRS Publication 15 (2026).
-
Checking Polish KSeF FA(3) invoices offline
einvoice-preflight now reads Poland's national FA(3) invoice format: NIP check digits, rate codes and the totals KSeF will reject.
-
Chip equivalence proofs are hard because of the design, not the rewrite
752 SAT proofs of ABC-rewritten netlists: proof time tracks multiplier width, and a classifier's high AUC is just design size.
-
A conformal shortlist for chip design search works on the mapped rung and fails on the generic-gate one
A conformal bound on a routed pilot keeps the best SKY130 design 99% of the time at about half the routed runs, but only if the cheap rung is a mapped netlist.
-
Counting drywall sheets by layout instead of area divided by 32
The wall-framing takeoff now counts drywall by laying out 4 ft courses, splitting them at doors and packing offcuts, instead of dividing square feet by a sheet.
-
A critical-path scheduler for a house build, running in your browser
The wall-framing takeoff has a build schedule card: paste activities, get the critical path, crew levelling and what a slipped inspection does to the finish.
-
CRM exports are messier than your test fixture: four fixes in a prospecting agent
A sales research agent that passed its tests on hand-built HubSpot data still misread Excel-style and Pipedrive-style files. What broke and what we changed.
-
Dating changes in a shallow clone takes 16 seconds, not 15 minutes
Our agents' friction report kept flagging already-fixed complaints because their shallow clones could not date file changes. A bounded fetch fixes it.