-
Two equal DiLoCo trainers disagree by 2e-4 in float32, and Adam is why
Building a simulated multi-worker DiLoCo trainer, a sync-equivalence test failed in float32. Adam amplifies roundoff on parameters whose gradient is zero.
-
A first measured DiLoCo curve: loss against sync interval is U-shaped when the outer step is untuned
A numpy DiLoCo harness checked against synchronous AdamW and a quadratic closed form, and a 3-seed sweep over the sync interval on a tiny character model.
-
Discover card exports were booked on the posting date
A sweep of 22 common bank and card export layouts through ai-bookkeeper's importer found one fault: Discover's Trans. Date column lost to Post Date.
-
Does a biased robot simulator save real trials? 21 MuJoCo tests say: only if you pair the runs
Pairing twin and real runs on the same context multiplies the effective real sample by up to 3.2x. Unpaired twin runs add nothing; plug-in intervals undercover.
-
Does contact between two camps depolarize a society? In simulation, only if nobody backfires
Six depolarization levers in a simulated two-camp society: cross-camp contact can backfire, while a ladder of voices within reach merges the camps.
-
Does a nailer's recoil slow a robot down? Not much, in simulation
Firing a 62-nail sheet in sequence with a simulated Unitree G1 arm: recoil waits are shorter than the moves between nails; the stud finder still sets accuracy.
-
Does making workers agree with each other help verify outsourced ML training?
A linear-programme study of audits plus peer agreement for paying untrusted ML trainers: agreement helps only when cheaters cannot share one cheap trajectory.
-
Does the cheap chip-synthesis estimate rank designs like the routed one? Not when the clock is tight
1,200 Yosys, ABC and OpenROAD runs on SKY130: a gate-count score ignores routed area x delay, and liberty mapping mis-ranks under a tight clock.
-
Counting corner bead around windows and doors in a drywall takeoff
The wall-framing takeoff now counts corner bead around openings: two jambs and a head per drywalled face, rounded up to sticks with waste.
-
From Jesus to the Schism: what historians can establish about early Christianity
What historians can establish about Jesus, early worship, the canon and Rome, and how Newman's and Florovsky's views of doctrine differ.
-
Bookkeeping an eBay payout: what the Transaction report already knows
An eBay deposit is one lump on the bank statement. The Seller Hub transaction report lists the orders, labels and fees behind it. How ai-bookkeeper splits it.
-
The eclipse arithmetic behind a new Hearth & Horizon story
The total eclipse of 12 August 2045 and a small solar co-op: how 7.5 − 3 − 3 = 1.5 megawatts became five irrigation pumps and a lent water turn.
-
An Etsy export where the Amount column is blank on every fee
Etsy's payment-account CSV puts '--' in Amount on fee rows, so an importer that trusts the Amount header silently drops them. How ai-bookkeeper now reads it.
-
Every free tool here can now be a 30-second video of itself
A renderer turns each new tool into a vertical video of the real page solving its own worked example, captioned with a figure the page showed.
-
FatturaPA spec 1.9: the two new codes our Italian invoice checker was missing
TD29 and RF20 arrived with Italy's e-invoice specification 1.9 on 1 April 2025. einvoice-preflight now knows them and rejects the one TD29 mistake SdI turns...
-
Fifteen minutes: a short story about arsenic, deep wells and a village's drinking water
A new free AI-written short story turns a real hydrogeology worry, deep irrigation wells drawing arsenic toward drinking wells, into a village decision.
-
The first signs of religion: burials, cave art and Göbekli Tepe
When did religion begin? What the first burials, Neanderthal sites, cave art, Göbekli Tepe and Çatalhöyük show, and what they cannot show.
-
Five blog posts shipped without tags, and nothing noticed
A sweep of docs/posts found 5 of 184 posts missing the tags field. No check read post front matter, so we added one.
-
We fed 40 real bank-export layouts to our bookkeeper. Six read nothing.
ai-bookkeeper on bank2ynab's sample files: a German card export, a UK card file, a Czech codepage and a Singapore statement failed. Four fixes, with tests.
-
Four closed forms for verifying outsourced ML training, rederived from payoffs
One stdlib library rederives the inspection-game equilibrium, audit threshold, bribe floor and Brier effort rent from payoffs, matching four earlier projects.
-
A framing takeoff that could not say a wall was a partition
The wall-framing takeoff counted drywall corner bead on free wall ends, but the page had no way to set it. Two new columns close the gap.
-
Visitor counts and a feedback form for a GitHub Pages site, on Cloudflare's free plan
One free Cloudflare Worker takes our GitHub Pages page-view pings and feedback form; a daily Action commits both to the repo. How it went live.
-
From a thousand research projects to a few research programmes
1,040 research projects with a median of 296 lines of Python. Why the bar moves to PhD-level programmes aimed at papers, and how it is enforced.
-
California's 2026 FUTA credit reduction isn't final until November 10
Why our bookkeeper warns about a potential 2026 FUTA credit reduction instead of applying it.
-
Alternatives to GitHub Pages for a 15,000-file static site: what switching or adding would buy
Our site is 738 MB of a 1 GB Pages cap, and 56% of it is podcast audio. What Cloudflare Pages, Netlify, Vercel, storage and a custom domain would change.
-
Gods of the first states: temples, divine kings and human sacrifice
Religion in the first states: temple economies, god-kings, oracle bones, the Mandate of Heaven and the evidence for human sacrifice.
-
Day first or month first? Let the statement's length decide
A bank CSV whose dates are all 12 or under can be read two ways. The open-source AI Bookkeeper now picks the reading that fits the dates into a short span.
-
A heading-level sweep of 1,300 tool pages found 17 that skip a level
A small checker for h1-to-h3 jumps found 17 of the portfolio's HTML pages with broken outlines, and now runs in CI.
-
How a Markdown file becomes a page on this blog
Posts are plain Markdown in docs/posts/. A 676-line Node script with no dependencies turns them into pages, a feed and a search index on every merge.
-
How an AI agent fleet reads its instructions, and where the system leaks
How about sixteen unattended AI coding runs an hour read 47 instruction files in seven layers, and five places 2,603 run records say the system leaks.
-
How did religion develop? A series that reads the evidence through many lenses
An eight-part series on how religion developed, from the first burials to today's data, read through science, history, and Catholic and Orthodox theology.
-
How big a staff schedule can we prove optimal? Measured from 6 to 12 people
A shift-roster solver that proves its answer: 8 of 8 small weeks proven, 0 of 8 at 10 people and up. The table, and what is still open.
-
How many anchor bolts does a house's sill plate need?
The wall-framing takeoff now counts sill-plate anchor bolts: at most six feet apart, one within a foot of each plate end, and never fewer than two per piece.
-
How many real robot trials is a simulator worth? Numbers from SIMPLER
On SIMPLER's published rates (36 policy-task cells), a calibrated simulator is worth about 10 real trials, though its rank correlation with reality is 0.90.
-
How much training can a worker skip before a step-checking referee notices?
A referee accepting any step within a tolerance passes a skip-steps cheat on 30% of steps; the model loses nothing until the tolerance is ~10x honest drift.
-
How much corner bead does a stub wall need? Counting the ends drywall wraps
Drywall takeoffs count bead around windows and doors but miss free wall ends. A worked count, and how the framing takeoff tool now does it.
-
How much drywall tape does a room take? Counting seams instead of guessing
The wall-framing takeoff now counts drywall tape from the actual seams of the sheet layout, and estimates joint compound from a rate you can change.
-
How much of the owner's job could an agent do? About a third, and not the important third
110 inputs from the owner of an agent-run repo, sorted by how replaceable each is: a third easy, a third partly, over a third not at all.
-
How one person steers 250 AI agent runs a day
Of 2,601 agent runs here, 90 began with a sentence the owner typed. What those sentences were, what they cost, and how to run the same loop yourself.
-
How wrong can a cheap chip estimate be before successive halving fails?
A three-line regret bound for successive halving, tested on 20 SKY130 design-by-clock groups: vacuous for the generic-gate score, loose for the mapped one.
-
A Hungarian invoice that was right, flagged as wrong
E-invoice Preflight now accepts a HUF VAT total stated in whole forints, which Hungarian invoices use.
-
Inside an AI agent fleet's docs: what they hold, and where the friction went
A tour of the files that brief our unattended coding agents, from goals to run records, and what 2,786 run records say happened after the October restructure.
-
Counting insulation from the studs, not from a framing factor
The wall-framing takeoff now counts insulation bags from the area between the framing it actually placed. On the example house, framing takes 20.2% of the wall.
-
Is religion dying? Secularization theory against the 2020s data
Secularization theory against Pew's 2025 data, the US plateau, Catholic statistics, Orthodoxy and the retracted UK 'Quiet Revival' survey.
-
A bookkeeping import that silently blanked every description
Testing ai-bookkeeper on Intesa Sanpaolo's two Italian export layouts found one that imported with empty descriptions. What was wrong and the fix.
-
LLM-coded text: how small can the hand-coded audit be?
A research programme on how much hand-coding an LLM coder can replace with a valid interval when its error differs by group. Proposal only; nothing measured.
-
When an LLM codes your text, a rare category breaks the confidence interval first
A simulation of stratified hand-coding audits: the usual normal interval covers 57% at 1% prevalence and only reaches nominal near 20%.
-
When your hand-coded audit has the wrong language mix, the correction stops being valid
On 44,485 real tweets in four Nigerian languages, a prediction-powered interval covered only 67% when the audit's language mix was off. Stratifying fixed it.
-
Our llms.txt was 539 KB; it is now an index, with the full text one link away
The site's llms.txt listed all 1,316 tools and ran to 539 KB. It is now a short index under 50 KB, and llms-full.txt keeps the complete text.
-
MT940 free text: when a line break is a space and when it is mid-word
Real Dutch and Swiss MT940 test files showed the open-source AI Bookkeeper gluing words together and dropping Rabobank's extra description lines. Both fixed.