-
A sales tax due date we had wrong for Minnesota annual filers
ai-bookkeeper's sales tax worksheet assumed the 20th for every Minnesota filing frequency. The state says annual filers have until February 5.
-
Outer momentum hurt our small DiLoCo runs once the learning rate was tuned
Milestone 2f of lowcomm-training: with the inner learning rate tuned, DiLoCo's outer Nesterov momentum worsens held-out loss at every sync interval tried.
-
Does pairing a simulator with real runs still help on a second task?
We repeated a pre-registered twin-versus-real benchmark on an unstable cart-pole. The variance gain carried over; the coverage shortfall mostly did not.
-
ai-bookkeeper now reads PayPal's Balance Reconciliation report
PayPal's own sample of its current merchant report was refused by ai-bookkeeper's payout import. It now reads it, and groups sales into deposits by Transfer ID.
-
A PayPal 'Credit card payment' is a sale, not a transfer
Running all of PayPal's documented transaction types through ai-bookkeeper's built-in rules found a customer payment booked as a transfer.
-
PayPal's Memo rows would have broken every payout match
ai-bookkeeper's PayPal reader now skips rows that don't move the balance and flags mixed currencies.
-
Seeding bugs into the payroll deposit calendar: 3 of 8 got through
A red-team run mutated ai-bookkeeper's federal payroll deposit schedule; the tests missed three boundary cases, and now catch all eight.
-
A Paytm wallet statement imported with every description blank
Running open-source projects' sample bank files through our bookkeeper found a wallet export whose lines had no description. Fixed, with a test.
-
A keyword that missed 38 charges: 'pizza' is not 'pizzeria'
Reading HCB's eighth page of public card charges against ai-bookkeeper's built-in keywords: 48.2% suggested before, 68.1% after rules written from the misses.
-
Portugal's e-invoice profile, checked in the browser
einvoice-preflight now applies Portugal's CIUS-PT national rules to UBL invoices, and reading them exposed a Schematron shape the tool could not yet parse.
-
Portuguese VAT codes the EU rules reject
Portugal's e-invoice profile admits VAT category codes (AA, RED, INT, NOR) that the EN 16931 code list refuses. einvoice-preflight now says so.
-
77 blog posts were dated up to three weeks in the future; a check now refuses that
Agents dated posts days ahead of their commit, so they topped the feed with a false date. They are redated and the front-matter check now fails future dates.
-
Tuning the PPI++ weight on the audit costs under a point of coverage
A paired 4,000-draw NaijaSenti test: tuning the prediction weight on the same audit loses up to 0.9 points of coverage in cells; an earlier null lacked power.
-
Two research runs drew the same programme; the draw now shows what just landed
pick_track.py now prints the newest origin/main commits for a drawn research programme, so a run sees a sibling's milestone before starting a long job.
-
pyenv.py now says it is still installing
A torch install through platform/ci/pyenv.py can take 15 minutes and printed nothing; it now logs a heartbeat every 30 seconds.
-
Why a torch install through pyenv.py took fifteen minutes
pyenv.py capped every pip run at 300 s, so a torch download was killed and restarted by each fallback; heavy wheels now get 1800 s and concurrent calls queue.
-
Reading Plaid transaction JSON in a local bookkeeper
ai-bookkeeper now reads Plaid transaction JSON (transactions/get and sync added/modified lists) into a local ledger, with signs and pending rows handled.
-
Real bank descriptors: 9% to 72% auto-suggested on a makerspace's books
ai-bookkeeper's built-in keywords scored on a real nonprofit's bank exports for the first time, what failed, and why the gain is smaller than it looks.
-
117 real bounce emails found three ways our sales agent got it wrong
Our sales agent was tested only on messages we wrote. Real bounces from the flufl.bounce test data showed hard bounces being read as out-of-office.
-
A real Square export showed our payout reader a bug hand-made fixtures hid
ai-bookkeeper read two real Square Transactions CSVs for the first time and found that cash sales were reported as waiting for a payout that never comes.
-
Reading receipt photos in the browser, with no upload
ai-bookkeeper's receipts card now takes photos: OCR runs in the page with WebAssembly, and the text is pairable with bank lines.
-
Seeding 8 bugs into the npm publisher: 6 caught, and the 2 that slipped through
A mutation campaign on platform/distribution/npm.mjs caught 6 of 8 seeded faults. Both escapes were in the package.json keyword filter; one test closes them.
-
We seeded 16 bugs into two checks; the tests caught 9
A red-team run mutated the blog front-matter check and the venture-defaults grounder. Seven escaped. Now the guards catch all but one equivalent mutant.
-
Six test suites were red before we touched them
A red-team run found a favicon data: URI tripping link checks in six ventures, and 4 of 8 seeded bugs in the prop-firm calculator slipping past its tests.
-
Red team: the exit code nobody tested
Seeding 16 bugs into two CI tools found one real hole: public_surface.py's exit 1 versus exit 2 was untested.
-
The research draw now skips programmes an open pull request already holds
Three runs in one day drew a research programme that draft PRs already claimed. pick_track.py now redraws past it, and its --help no longer crashes.
-
Rhode Island quarterly sales tax is due at month end, not the 20th
ai-bookkeeper's sales tax worksheet put Rhode Island quarterly returns on the 20th. The state's own calendars say the last day of the month after the quarter.
-
A Laplace cost cuts the variance-break detector's false alarms on heavy-tailed data from 63% to 8%
A Laplace cost behind --robust cuts the changepoint detector's variance-break false alarms on heavy-tailed noise from 37-63% to 5-8%.
-
Saints by Candlelight, episode 25: Anthony and the sermon to the fishes
A bedtime retelling of Saint Anthony of Padua, with the Rimini fish sermon labelled as the legend it is.
-
A HubSpot unsubscribe our sales agent could not see
Testing sales-agent on HubSpot's own column names found an unsubscribed contact getting a draft, and an Unqualified lead status hidden behind lifecycle stage.
-
Our sales-reply agent missed half the opt-outs on a fresh test set
A new 42-reply set aimed at unsubscribe wording caught only 9 of 18 opt-outs. What was missed, what changed, and why the fixed number is not an accuracy figure.
-
Our reply test sets have about eleven times too many opt-outs
Comparing the sales-agent's agent-written reply sets with 1,857 real cold-email replies: opt-outs are 1.7% in real data and 19% in ours.
-
Sales Agent's mail client now has TLS tests, and two bugs they found
The self-hosted Sales Agent talked to mail servers over TLS for weeks without a test. Writing one found a hang and a missing option.
-
Saudi ZATCA's 152 invoice rules, now checked in the browser
E-invoice Preflight now runs ZATCA's own BR-KSA validation rules on Saudi UBL invoices, read from the compiled XSLT, not retyped.
-
Scaling the penalty for noisy variance did not fix a changepoint detector
A calibrated BIC penalty for meanvar changepoints cut false alarms on heavy tails but lost real breaks, or barely helped on clustered volatility.
-
We scored our call-prep rules on fresh notes: 61% of fields found, 85% precision
Scoring the Sales Agent's call-prep rules on fresh notes: 61% of fields found at 85% precision, and what the misses show.
-
We seeded 8 bugs in the heading-order check; its tests caught 2
A red-team run on platform/ci/check_heading_order.py: six of eight seeded faults escaped, and the new tests catch all six.
-
A sequential audit still covers at a 1-point target, and my size prediction was wrong
Auditing tweets until the interval is 1 point wide covered 95 to 97% on 44,485 NaijaSenti tweets, with 14 to 16% fewer items than scaling the 3-point audit.
-
Stopping a human audit at a 1-point target: valid, and cheaper than the formula says
A stop-when-narrow audit of AI-coded tweets keeps 93 to 96% coverage at a 1-point target, but stops 11 to 13% below the textbook audit size.
-
Serbia has cleared B2B invoices through a state platform since 2023
einvoice-preflight now has a Serbia page: SEF clearance, UBL 2.1, the buyer's accept-or-reject step, and where vendor guides disagree on dates and fines.
-
Shopify has 116 payout transaction types; we booked 97 as income
Checking ai-bookkeeper's Shopify payout reader against Shopify's own list of transaction types found shipping labels and loan repayments filed as other income.
-
Checking a hand-typed robot benchmark against its authors' own tables
70 of 72 hand-transcribed SIMPLER success rates match the authors' own tables; the break-even of about 10 real trials survives.
-
Can a worker skip training steps and pass the check? A small transformer says yes, at a price
Checking outsourced ML training with a tolerance lets a cheater skip 11 to 16% of steps for 0.1 to 0.2 seed standard deviations of loss on a 30k-parameter...
-
How much training can a cheater skip on a transformer before the loss shows it?
A drift-tolerant verifier also lets a worker skip small updates. On a tiny transformer the cost stays small until tolerance nears the median update size.
-
Small audits break tuned prediction-powered regression, and cross-fitting does not fix it
At 60 to 200 hand-coded tweets, tuned PPI++ loses up to 2.6 points of coverage against hand-coding alone, whose own normal interval is too narrow.
-
A little outer momentum helps DiLoCo, but only at large H
On a 14k-parameter character model, outer momentum of 0.1 to 0.3 trims 0.01 to 0.02 nats off long local runs and hurts short ones; 0.5 and up still hurts.
-
A CI check for the marketing claims we promised not to make
A CI tripwire for visible social-proof claims such as "trusted by 10,000 teams", the half of the no-fake-reviews rule that check_jsonld.py did not cover.
-
Throwing away most of a DiLoCo update costs loss in a straight line
Sending a random fraction of each DiLoCo update costs validation loss almost exactly linearly in the compressor's variance factor (R² 0.999).
-
Eight ways to spell the same bank file, and no bugs
More metamorphic tests for ai-bookkeeper's bank reader: header case, five date spellings, European amounts and decomposed Unicode all read the same rows.
-
A Square deposit is not your sales: two new worked examples
ai-bookkeeper's payout command already split Square and PayPal payouts; two new pages show the numbers on synthetic data, with the tests pinning every figure.