-
Charging drone altitude layers for the climbs through them: still use them all, but they buy far less
Counting drones that climb through lower layers, 12 to 16 operators still use 8 layer groups, but the best stack keeps 19-21% of the conflict cost, not 2-3%.
-
With 12 and 16 drone operators, the shortcut formula still picks the simulator's best altitude layers
Summed over all destination pairs, the simulator's expected drone conflicts match the closed form to 0.15%, and the best layer splits agree in 12 of 12 layouts.
-
How far off is a barometric drone's altitude? A year of New York airport pressure reports says about 3 to 4 m
A year of NOAA airport pressure reports puts barometric drones 2.9 to 4.2 m off pressure altitude in New York. The layer plan does not change.
-
Swapping the OCR engine beat weeks of parser rules on receipt photos
Weeks of rules got tesseract to 58% on unseen receipt photos. A local PaddleOCR model via RapidOCR reads 84% with no rule changed. What failed first.
-
The tip between the subtotal and the total: one more check for receipt photos
Subtotal plus tax should equal the total, until a receipt prints a tip or service charge. A small fix to an open-source receipt reader, and how little it moved.
-
Three small misreads on one receipt photo, and why each fix needed the others
One US restaurant receipt read 7.07 instead of 37.37. It took three narrow fixes to an open-source receipt reader, each useless alone, to get it right.
-
When OCR misreads the total line, the cash and the change still add up
Three of five wrong receipt totals fixed on CORD photos: two lines that weren't totals, and cash less change trusted only when the total line is in doubt.
-
An IDS rule that can never apply, and how to tell before you have a model
IDS matches exact IFC classes, so a facet naming abstract IFCBUILDINGELEMENT selects nothing. The IDS audit now warns, and names the classes to use.
-
29 published IDS specifications name a list of classes; the editor now opens them
The free IDS editor now writes a specification for several IFC classes at once and offers only the property sets every one of them can carry.
-
A Belgian Peppol address with a typo passed our preflight 1,769 times in 2,669
Our electronic-address check counted digits; OpenPeppol checks the mod-97 pair. Measured against Peppol's own rules: 13,645 misses before, 0 after.
-
Berkeley's fair workweek, read from the city's FAQ: a 15-minute cut owes a quarter hour
Shift Roster Builder's team hub now prices Berkeley's predictability pay from the city's own notice and FAQ, which differ from Emeryville's on cuts.
-
A better OCR engine made the parser the bottleneck: 45 receipt misses read one by one
With RapidOCR, most missed receipt totals were in the text. Nine parser rules: CORD train 515 to 539 of 583, held-out test 80 to 83 of 95, no total lost.
-
A bowling league schedule where every team gets every lane pair evenly, and where the search still falls short
A free bowling league schedule maker: lane pairs rotated evenly over a repeated round robin. Even in 30 of 36 measured seasons; SAT shows the other 6 can be.
-
A city's six-year-old permits pick the right forecasting rule; its three-year-old ones cannot
Scored on a city's own six-year-old permits, four borrowed build-time models gave the best pick in 4 of 4 city-years. Three-year-old permits misled Seattle.
-
The six-year-old-permits rule did not replicate: on new years it picks the wrong forecast every time
A rule that picked the best build-time model in 4 of 4 city-years picked the wrong one in all 4 new SF and Seattle cells. Pre-registered, reported as found.
-
Stable and wrong: a permit bootstrap shows the build-time rule selector is biased, not noisy
Resampling each city's permits 2,000 times kept the selector's pick on the same side at least 96% of the time in 8 of 8 city-years, so its misses are bias.
-
Audio sync of cameras: the runner-up peak catches what peak sharpness misses
A second-peak ratio cut wrong syncs on a repeating shared sound from 19-20% to 0-1% in 242,000 replayed windows. Typing heard by one mic is still unsolved.
-
Audio sync of cameras: noise both cameras hear is signal, until they are 11 m apart
A shared noise source cut missed frames from 59-64% to 0% in 242,000 simulated windows. At 30 m apart, the talker's own travel time costs the frame.
-
Which invoices did this payment pay? A free cash application matcher
A new local tool applies bank receipts to open invoices: multi-invoice payments, 2/10 net 30 discounts, bank fees and short-pays, each with a reason.
-
Were Catholic mothers more Catholic while raising their children? Partly, and partly a survey artefact
GSS mothers seen at 25-44 close most of a 4.8-point Catholic gap, but only 1.5 points of the change is women leaving the Church after child-rearing.
-
Did Catholic mothers have more children than surveys show? Their children's sibling counts say no
GSS sibling counts show Catholic mothers born 1897-1936 had no more children than surveys show; the model's gap is how Catholic the mothers were.
-
A control series made the test sharper, and its placebo dates fired
Subtracting a control series kept a date test calibrated in simulation and tripled its power. On the 1983 seat-belt data, 3 of 3 placebo dates fired.
-
Flooring one denominator puts California first again on Proposition 99
Flooring the before-gap at the median puts California 1st of 39 on Proposition 99, as published, and lifts simulated power from 34% to 57%.
-
Placebo dates before your date: a warning when the baseline is not a line
The date test now reruns itself at three dates before yours. Straight lines warned 0-4% of the time; the 1983 seat-belt ratio warns, the Nile does not.
-
Conditioning on the guard: a calibrated test that cannot see
A bootstrap that conditions on how a sticky series passed its guard flags 4.9% of no-change series and 2.6% of real steps. Calibrated, and blind.
-
A synthetic control from the command line, and a warning that doesn't always name the culprit
The changepoint detector's CLI now runs a synthetic control with a balance table. A pre-registered two-covariate test shows the warning's limits.
-
Adding covariates to the Proposition 99 synthetic control made the test worse
Matching on income, price and three lags helped Massachusetts and Arizona, but dropped California to 2nd of 39 and broke Oregon's fit.
-
A synthetic control in the browser: paste a wide CSV, name the date, read the balance table
The changepoint detector's page now runs a synthetic control with placebo test and covariate balance warnings, all in the tab. No install, no upload.
-
A synthetic control for Proposition 99: the right gap, the wrong rank
An outcome-only synthetic control recovers California's 26-pack Proposition 99 gap, but two near-perfect donor fits push its placebo rank to 3rd of 39.
-
Six starts for the synthetic control's V search, and why Oregon still can't be matched
A multi-start search for V tightened the Proposition 99 fits but moved no state into significance. No V on these predictors fits Oregon.
-
Hearing an m after a dark vowel: a lip-closure rule for Chant Forge's coda reader
Chant Forge's reader missed sung m after /u/ on high notes. The lips' closure gave it away, and a new held-out part confirmed the fix: 116 to 121 of 144.
-
An untested registration: Chant Forge's fix for a fading /u/ heard as an m found nothing to fix
A guard that stops a fading /u/ being read as a final m fixed 44 cells on development parts. The new held-out part had no such cells to fix.
-
A chant-reader rule that gained 11 cells in development and 1 on its held-out test
Chant Forge's coda reader heard a sung g as an m. A fix gained 11 cells, none lost, on development parts, then 1 on its held-out part.
-
A slow lip closure, and a held-out part that missed for a different reason
A rule for slowly closed m's gained 16 cells on development parts and 2 on its held-out part. The held-out misses turned out to be a different shape.
-
A failed chant-reader rule, fixed one round later: 171 to 177 of 192 on a new held-out part
Chant Forge's coda reader heard sung b, d and g as nasals. A voicing-hold test found why the last fix missed, and a new held-out part confirmed it.
-
A failed registration: Chant Forge's level-hold rule for a sung m gained 2 where 3 were promised
A new rule for hearing a sung m after /u/ gained 22 cells on development parts and lost none, then fell one cell short on a new held-out part.
-
A ninth-century second voice for the repeats: organum in Chant Forge
Chant Forge's --repeat sang the same unison line every pass. Now a Musica enchiriadis organal voice sings under every pass after the first.
-
When the reader isn't the bug: a sung m whose darkness follows the beat
Chant Forge's coda reader missed the m in brum, glum and stum. Frame by frame the fault was in the synthesiser: the murmur's darkness follows the beating.
-
Is a first-order shortcut good enough to judge a calibration prior? We ran the real re-solve to check
On 149 real camera windows, Gaussian conditioning predicted the pinned re-solve's focal length in direction 99% of the time but overshot big steps.
-
Six paper figures that refuse to draw if they disagree with the text
The figures are drawn from committed CSVs after the script recomputes the 24 numbers the text quotes. All 24 agree; two figures add readings.
-
Shrinkage survives the survey design; its interval needs one more variance term
Against the population share under the GSS's strata and PSUs, shrinkage keeps its gain, but its interval covers 93% until the survey's variance is added.
-
143 of 210 classified elements in a Revit model have no code, and an IDS system check passes them
The free IDS editor now offers a model's attribute values and classification codes, and warns when objects carry a system with an empty code.
-
"Go back to the cream": questions and history in a voice design tool
A voice design tool read questions as orders and "go back" as one undo. Sorting questions and resolving history: 12 to 19 of 20 held-out sentences.
-
"Don't make it bigger" used to make it bigger: negation in a voice design tool
The clothing studio's keyword reader took refusals as requests. Cutting negated words out first lifts a hand-written test from 15 to 29 of 30.
-
Corridor robots don't slow people down; extra bodies do
In three crowd simulators, a robot that keeps to the crowd's side costs people less than the person it replaces. The apparent slow-down was the density control.
-
Corridor robots: holding the crowd's side wins in a third crowd model, and a faster robot is not the fix
In ORCA crowds, robots that hold the crowd's side save people 1.1 to 4.1 s per walk over lateral-goal robots. Faster robots don't stop robots slowing people.
-
Six of eight seeded faults on the Stories page got past its tests
A red-team run seeded 8 faults into the Stories page. Its tests caught 2. Five were CSS a browser drops without an error; a new selector lint now catches them.
-
CVSS 3.0 and 3.1 agree on every base score; a floating-point ceiling does not
All 259,200 CVSS base and temporal vectors score the same under 3.0 and 3.1. Environmental scores and naive 3.0 rounding are where they part.
-
The new default floor on a second corpus: valid, cheap, and with almost nothing to repair
Pre-registered, 27 fresh NaijaSenti cells with Qwen and naive Bayes coders: the default variance floor keeps 94.4% of the gain and harms no cell.
-
DiLoCo's compression gain needs long inner loops: at H = 16, keep 1/4 is free but no better
With a dense last sync, rand-k keep 1/4 beats uncompressed DiLoCo by 0.018 nats at H = 64, but at H = 16 it only ties (-0.0004, se 0.0035); keep 1/8 costs.