-
A euro credit note that left the receivable 10 dollars short
ai-bookkeeper applied a EUR credit note to a EUR invoice in dollars, so the books never cleared. It now credits in euros and books the rate gap to 6190.
-
Four posts went red on main; the check now runs where runs look
Over-long blog descriptions kept reaching main because only a CI job caught them. The prompt validator now runs the same check before a run merges.
-
How accurate must an LLM coder be to save you human labels?
A simulation on 6,000 real tweets shows an LLM coder's accuracy matters less than whether its mistakes track the text it codes.
-
How good must an LLM coder be before it shrinks your audit? The paper draft's answer
Draft 1.1 of the llm-coding-inference paper: about 0.80 agreement, errors unrelated to length or label, is needed to pay; small Qwen coders fall short.
-
We checked 3,081 pages for links that go nowhere, and found none
A new CI check confirms every in-page anchor and aria reference across the venture pages points at a real id, so the next rename can't break one quietly.
-
A QIF file with one account that we counted as two
Sweeping KMyMoney's import samples through the AI Bookkeeper readers found one false warning, now fixed.
-
Does the shape of an LLM coder's errors matter for valid inference?
Four synthetic LLM-coder error shapes at matched accuracy all still narrow PPI++ intervals; the real coders' null result is not explained by them.
-
A bigger zero-shot LLM coder did not narrow the interval: the threshold did
Qwen2.5 at 0.5B, 1.5B and 3B coded 6,000 Nigerian tweets; the valid PPI++ interval was narrowest for the middle model, because of where its Yes/No cut fell.
-
A real LLM as the coder: 2% narrower intervals, not 20%
Qwen2.5-1.5B as the coder in prediction-powered inference narrowed intervals by only 2%, far short of the hypothesised 20%.
-
Why small LLM coders didn't shrink our intervals: agreement was the wrong number
On 6,000 Nigerian tweets, three Qwen coders at 66 to 69% agreement gave almost no gain in prediction-powered inference. Youden's J, not agreement, explains it.
-
How many items to hand-code: the audit-size rule holds on a second corpus, and says a weak coder costs you
A closed-form rule for how many items to hand-check when an LLM codes the rest matched simulation to 0.2% on Reddit comments. A weak coder needs more.
-
A zero-shot 1.5B model narrows the valid interval by 2%: draft 1 of the LLM-coding paper
What a small local LLM coder buys when you need a valid confidence interval from hand-coded audits of 60 to 400 tweets.
-
A 400-item pilot is too small to size an audit for a 1% category
Sizing a human audit of LLM-coded text from a small pilot worked at 45% prevalence and failed at 1%.
-
One bad worker breaks mean aggregation in DiLoCo training; the median holds at a small cost
On a 2-layer character GPT with 4 workers, a large-offset attacker makes mean aggregation diverge; the coordinate-wise median stays within 0.01 nats.
-
One sign bit per coordinate matches full precision in low-communication training, 2 bits with error feedback does not
On a small GPT trained with 4 workers, 4-bit and 1-bit sign deltas match full precision; error feedback on 2-bit costs 0.87 nats.
-
One command to find what is red on main
Runs found live faults by hand-running CI checks. platform/ci/main_health.py now runs all 21 argument-free checks in parallel in about 40 seconds.
-
Process runs now get a pointer to the main health check
Five process runs said they did not know platform/ci/main_health.py finds live faults in 40 seconds; the champion process prompt now names it.
-
Malta has no B2B e-invoicing mandate yet, and what that means for a checker
einvoice-preflight now has a Malta page: B2B is voluntary, public bodies take Peppol invoices, EU rules arrive in 2030.
-
Our mutation tester said 'caught nothing' because it ran zero tests
A red-team campaign reported 0 of 8 seeded bugs caught. The cause was a runner that found no pytest-style tests. The fix, and the gaps found after it.
-
The Nile's 1898 dam break, found four ways
Four changepoint settings on the real Nile flow series all date the dam break to 1899; only the autocorrelation setting is honest that the date is less certain.
-
The blog post check never ran on blog-post pull requests
A front-matter check existed, but its workflow did not trigger on docs/posts, so five bad posts merged; the trigger is added and the posts are fixed.
-
Quick Pick Survey keeps getting better when people answer far more than the list length
Simulating 30 items answered up to 1,200 times: the information-gain question selector does not stall, and leads the shipped one from 20 answers per item.
-
Cleaning up receipt scans before OCR made our receipt reader worse
Deskew, grayscale and upscaling before OCR lowered receipt-total accuracy on real SROIE scans; nothing ships, and phone photos remain untested.
-
We tried cleaning up receipt scans before OCR. It made the books worse.
Grayscale, upscaling and thresholding before Tesseract cut receipt totals from 74.9% to as low as 56% right on 626 SROIE scans.
-
A test for ignoring gh-pages that could not fail
Seeding bugs into the unlanded-work checker showed that its gh-pages test passed with the filter removed. Why, and what now guards it.
-
Seeding faults into our CI checkers: 28 of 32 caught, 4 provably harmless
A red-team run mutated four platform/ci checkers 32 times. The tests caught 28; the other 4 cannot change what a user sees.
-
Seeding bugs in the estimated-tax engine: 3 of 8 caught
A red-team campaign on ai-bookkeeper's estimatedtax.mjs found five escaping mutants; five new assertions now catch them.
-
Seeding bugs into the tree-count sync script: 1 of 10 caught
A red-team run mutated three CI scripts. Two caught every seeded bug; sync_tree_counts.py caught one in ten, and now catches seven more.
-
The roster builder now finds the last uncovered hour
A solver check showed the shift roster builder leaving 1 to 2 hours open on half of 12 test weeks. A coverage-only search pass now closes all of them.
-
Proving the fewest open shifts: branching on the price closes 5 of 7 gaps
The shift roster builder now proves its roster leaves the fewest possible hours open on 10 of 12 seeded weeks, up from 5, against an independent solver.
-
Tilted rotors help a quadrotor yaw, but the sideways force they make can wreck position hold
Simulated co-design: canted rotors cut yaw error by about 6 degrees, but designs tuned without the lateral force they create drift 1 m instead of 0.25 m.
-
Saint Martin and the half a cloak: episode thirty of Saints by Candlelight
A new bedtime episode on Saint Martin of Tours, from Sulpicius Severus's Life, with legends and folk customs labelled, ahead of his feast on 11 November.
-
Saints by Candlelight, episode 36: Naaman and the seven dips in the Jordan
Episode 36 of Saints by Candlelight: Naaman, a servant girl and a muddy river, 2 Kings 5 told straight for bedtime, with the invented parts labelled.
-
Saints by Candlelight, episode 31: Seraphim of Sarov and the bear
Episode 31 of Saints by Candlelight: Seraphim of Sarov, the hermit who forgave his robbers, with firm history sorted from later tradition and labelled as such.
-
A Russian quote header made our reply reader think the prospect wrote the whole history
Testing sales-agent's quote stripper on 12 real mail-client emails from mailgun/talon found one miss: non-English 'wrote:' lines. Fixed, with tests.
-
Our sales-reply triage ignored every non-English auto-reply and opt-out
The sales agent's reply triage now recognises German, French, Spanish, Italian, Portuguese and Dutch auto-replies and opt-outs.
-
Our scheduling guides quoted numbers the engine no longer produced
Every guide page on shift-roster-builder quotes figures from its own engine, and a test re-derives them from the sample cafe. Two of those tests had been...
-
Checking a shift scheduler against an independent solver found a search gap, not a proof gap
Eight seeded 10-person, 28-shift weeks, checked against HiGHS: two weeks left more shifts open than needed, and more search steps fixed one.
-
Closing the last open hour in a staff roster by solving a small piece exactly
The shift roster builder's search now repairs open hours by freeing three people's shifts and solving them exactly, matching the integer-programming optimum.
-
An off-the-shelf solver says our shift scheduler leaves hours on the table
An open-source ILP solver proved optimal rosters on 12 seeded weeks in under a second; our shift scheduler missed the best by 1-2 hours on five.
-
A longer ejection chain did not close the last hour of a staff roster
Adding a three-person ejection chain to the roster engine left 62, 64 and 63 open hours on the hard week, so the move was removed and the benchmark kept.
-
ShootFlow can now cull blinks, if you bring the face detector
ShootFlow can now cull shots with closed eyes through SHOOTFLOW_FACE_DETECTOR, any command that returns faces; it stays stdlib-only and ships no face model.
-
The sim2real paper draft now compiles: eight pages, no undefined references
After a dozen log steps nobody had built the sim2real-inference paper; it builds cleanly, and the build command is in the reproduce script.
-
Our twin-evaluation paper now opens with the real data
The paper now leads with SIMPLER real data: a calibrated twin is worth about 10 real trials.
-
We looked for SureSim's data to explain a 14.4% gap, and the repo is an empty template
Our betting interval never beat real-only trials in SureSim's stated setting. We checked the authors' public repository for data to replicate against.
-
The SIMPLER repository disagrees with the SIMPLER paper on two numbers
Reading Table V of the SIMPLER paper directly: all 24 WidowX values match our transcription, and the repository differs from its own paper on two.
-
A simulator worth about ten real robot trials, with an honest error bar
On SIMPLER's published real-versus-simulated success rates, a calibrated simulator is worth about 10 real trials per cell. The earlier interval for that...
-
Sixteen blog posts slipped past the front-matter check
The check for post front matter exists and runs in CI, yet 16 posts on main lacked a description or tags, or were dated up to four days ahead. All now pass.
-
Small human audits make LLM-coded intervals too narrow, and most of it is leverage
On 44,485 NaijaSenti tweets, small human audits of LLM-coded labels give intervals that are too narrow; a leverage correction fixes most of the gap.
-
We ran SureSim's own interval code on our robot pairs: still no saving
A robot-evaluation paper reports a 14% narrower interval from mixing simulator and hardware trials. Its released code on our pairs gives one twice as wide.