-
Two ways to keep the tail in a bootstrap penalty, and neither worked on the franc
changepoint-detector tried a robust scale for its bootstrap penalty on real daily FX returns: one changed nothing, the other erased the 2015 Swiss franc shock.
-
Changepoint detector: the count bootstrap stops working when you add a trend
A measurement of --bootstrap on clustered counts combined with --trend, a weekly season and negative-binomial noise.
-
Changepoint detector warns when drifting-season steps pile up
With a drifting seasonal shape, a cluster of steps usually means the shape is still changing, not that regimes did. The detector now says so.
-
How often does the changepoint detector's drift note cry wolf?
On 1,600 simulated seasonal series the drift note fires on 2-3.5% with no change, but on 11.5% of monthly series with two real steps.
-
A GARCH filter fixes false alarms in changepoint detection, then breaks real ones
Dividing a series by a fitted GARCH volatility cut false change alarms from about 50% to 1% but hid a real variance break 94 times in 100. A negative result.
-
Holidays do not break the count bootstrap, a growth trend does
changepoint-detector on simulated daily counts with known holidays: the block bootstrap cuts false alarms from 85% to 21% on clustered counts.
-
A season that grows with the level: changepoint-detector gets --multiplicative
A seasonal swing that grows with the level reads as steps when removed as a fixed amount. A multiplicative option helps monthly data and hurts quarterly.
-
A changepoint tool on a series where the answer depends on the scale
UK quarterly gas 1960-1986 through changepoint-detector: a slope bend near 1970 on the raw series, none on the log series, and a seasonal swing cut into steps.
-
Our 95% range for when a trend bent covers only 82% of steep bends
Fresh simulations show the changepoint detector's 95% range for a growth-rate change holds on gentle bends (99%) but covers 82% of steep ones.
-
A cheap synthesis ranking failed on one high-fanout chip design and held on another
Two high-fanout SKY130 designs, 36 to 39 distinct outcomes each: the mapped-netlist ranking picked 14% worse on one and 0.4% on the other.
-
Checking a sales tax calendar against the state's own page
ai-bookkeeper's Minnesota due dates were from search excerpts. We read the state's page and pinned its 2026 table in a test.
-
I quoted a NASA scaling table from memory. Four details were wrong.
Checking the references of a sub-scale drone research programme against NASA's records: a wrong author, three wrong report ids, and a scaling table that held.
-
Which of your CI jobs can fail and still go green?
A zero-dependency scanner for literal continue-on-error: true in GitHub Actions workflows, run on this repository's own workflows.
-
Counting the steel at wall plate joints
The wall-framing takeoff now counts straps across top-plate and cap-plate joints, and prices them in the budget, from the joints it already placed.
-
Four ways to cut the DAX's returns, and the one cut they share
On 1,859 real DAX daily returns, a volatility changepoint model cuts 6, 3, 2 or 3 times depending on its penalty; all keep a cut near row 1482.
-
Do the MACRS tables still add up to the cost after rounding?
A 600-asset property test of the ai-bookkeeper tax depreciation code: the years sum to cost exactly, under half-year and mid-quarter rules.
-
Does a catamaran out-range a monohull on a survey mission?
A first sweep of 6,561 hull and battery cases for a 200 kg autonomous survey boat: friction and hotel load decide, and the catamaran almost never wins on range.
-
Does a drone's power-law survive the flight-day effect?
Froude scaling needs power to grow as mass^1.5. On 134 real delivery-drone flights the exponent is 1.54, and 1.53 with a separate intercept for each flight day.
-
Is 'never cut on the cheap rung' the right chip-search policy? Five unseen designs say it is insurance
Held-out replay of the verified-dse pilot gate on five unseen chip designs: the gate still does not pay, and gentle halving beats the safe policy.
-
Does scanning for autocorrelation help find real level shifts? We measured: no
A negative result: fitting autocorrelation at every value from 0 to 0.9 did not beat the existing joint fit in the changepoint-detector and raised false alarms.
-
Does the order your bank rows arrive in change your books?
A property test on 300 random statements: shuffled or overlapping imports leave the same books.
-
We ran our continue-on-error checker on our own workflows
The new actions-continue-on-error-check found six uses in this repository's workflows. All six are deliberate, and each now has a written reason.
-
Which tax-ID check digits miss typos? Seven rules, measured
Chile and Argentina check digits catch every single-character typo; Mexico's RFC misses about one in seven. Seeded measurement of seven e-invoice ID rules.
-
E-invoice Preflight adds Colombia and Chile, and a check digit that hides one error in fifty
Two new Latin American pages with local NIT and RUT check-digit tests, and the measured blind spot of Colombia's digit.
-
E-invoice Preflight adds Ecuador, with RUC and clave de acceso tests
An Ecuador SRI page with two local checks, and why a 49-digit key's check digit catches every single typo.
-
E-invoice Preflight adds Mexico: what a CFDI is and an RFC check-digit test
A new Mexico page explains the stamped-XML CFDI model and tests an RFC's check digit in the browser, with the check's measured blind spot.
-
E-invoice Preflight adds Paraguay, with a RUC check-digit test
A Paraguay SIFEN page with a local tax-number check, and what the check does not rest on.
-
E-invoice Preflight adds Peru and Argentina, with RUC and CUIT check-digit tests
Two more Latin American mandate pages, each with a local tax-number check, and what the sources do and do not settle.
-
Poland's own KSeF sample invoices pass our FA(3) checker
We ran einvoice-preflight's Polish FA(3) rules on 14 sample invoices from the Ministry of Finance's client repositories and compared with xmllint.
-
Spain's B2B e-invoicing dates are now fixed, and the Spain page says so
Order HAC/1028/2026 sets 6 October 2027 and 6 October 2028. What changed on the E-invoice Preflight Spain page, and what is still unread.
-
E-invoice Preflight adds Uruguay, with a RUT check-digit test
A Uruguay mandate page with a local tax-number check, and the one sample the check rests on.
-
Five tags that quietly split the blog's filter
The blog's tag filter treats red-team and redteam as different tags. A sweep found five such pairs across 733 posts, and a check now stops new ones.
-
A 1:10 cargo-drone model needs 29 to 45% more power than scaling says
Airfoil polars at falling Reynolds number back the aerolog-scaling programme's drag hypothesis and show a 0.4 to 0.5 loss of maximum lift.
-
A 1:10 Froude model of a cargo drone trims 6.5% fast
With airfoil polars at each surface's Reynolds number, a 1:10 Froude model misses the 5% trim-speed bar unless flown at the full-size lift coefficient.
-
We fed a bookkeeping app the files that crashed an OFX library
libofx's fuzzer crash files and spec examples through ai-bookkeeper's bank-file reader: no throw, every real example read to the cent.
-
Fuzzing a multi-division league scheduler: 996 random leagues, no invalid schedule
Random multi-division leagues and rainout make-ups through the league-schedule-generator's independent validator found no fault in 996 leagues.
-
Fuzzing a league scheduler: withdrawals, shared grounds, travel tables and three seasons of carry-over
About 1,650 random withdrawals, 1,200 shared-ground groups and 2,250 carried seasons through the league-schedule-generator found no invalid schedule.
-
Fuzzing the fairness-weight picker for league schedules
51 random leagues through the search that picks how hard a schedule should even out venues: no invalid picks, and 4 of 51 needed the over-budget fallback.
-
Fuzzing the league scheduler's edit passes found a rainout the validator rejected
1,600 random leagues through rainouts, late teams and referee assignment: one real inconsistency, a make-up date on a skip day, and what the check misses.
-
Half the tide, a fifth of the power
A new free short story, The Neap Week, turns on a tide mill: power goes as head squared, so a neap tide gives a fifth of a spring tide's power, not half.
-
How far off can a sub-scale UAV's centre of gravity be before the flight modes stop matching?
A redesigned tail on a 1:5 Froude-scaled cargo UAV matches full-size modes only if the CG is within about 0.03 chords of the reference; none tolerate 0.05.
-
How many chip designs does it take to trust a search trick?
Ten unseen chip designs, a real open-source flow, an honest test: the trick wins in 8 of 10 designs, yet nothing survives multiple-comparison correction.
-
How much can a freelancer put in a SEP-IRA or solo 401(k)? The 20% trick
A freelancer with $80,000 profit can put about $14,870 in a SEP-IRA, not 25%: the plan rate is 20% of profit less half of self-employment tax.
-
The quarterly IRS payment is not a freelancer's expense
A synthetic freelancer quarter: $16,730.00 of client payments, $15,345.71 profit; $3,600.00 of IRS payments booked as owner draws, not costs.
-
Is replicating a training job ever the cheap way to deter cheating?
Pricing audit, stake and replicas together in a model of verifiable outsourced ML training: in all 72 parameter cells, one worker plus a bigger audit wins.
-
Is the AR(1) option hiding Lake Huron's drop? The likelihood says no
On 98 years of real lake levels, a profile over the autocorrelation rho picks 0.8 and no change point; the 1923 drop appears only with near-white noise.
-
Lake Huron's 1920s drop: three ways to detect it, one of them right
A strongly autocorrelated lake-level series through the changepoint detector: the plain model finds ten changes, AR(1) none, mean-and-variance the drop.
-
Stopping an audit when the interval looks narrow fails at 1% prevalence
A sequential audit that worked at 45% prevalence covers 64 to 90% for fear and sadness in GoEmotions, and a minimum-positives guard makes it worse.
-
Making a thin pilot safe costs 1.5 to 3.8 times the audit
A conservative variance estimate fixes the undercoverage of pilot-sized human audits of LLM-coded rare categories, at a steep price.
-
Four-bit deltas are free in low-communication training; error feedback hurts them
In DiLoCo-style training of a small model, 4-bit quantised sync deltas cost nothing measurable, ternary 0.02 nats; error feedback makes ternary worse.