-
The demolition finding was a censoring artefact: a fold rule fixed in advance helps Chicago, hurts Austin
Seattle's short 2024 demolition build tail is right-censoring at the pull. A class-merge rule fixed in advance helps a borrowed model in Chicago, hurts Austin.
-
Why a work-type label hurt one permit forecast: demolition, not new construction
A work-type label hurt a cross-city forecast of Seattle build times in 2024. The cost sits in demolition permits, and it did not repeat in 2023.
-
Stop naming demolition: one class merge removes two thirds of a borrowed permit model's build-time error
Folding demolition into alteration cuts a transferred permit model's Seattle build-time q0.9 cost from 3.0 to 1.0 days and keeps its issue-time gains.
-
Score build-time tails on labels you can see: a complete-case score reversed our result
On Seattle permits seen in full, a merge that looked like a 2-day build-time q0.9 gain is a 1.7 to 2.5-day loss; IPCW recovers only 10 to 40% of the gap.
-
Austin's 15-day build-time penalty came from 649 Seattle permits
A work-type class helps San Francisco's borrowed build forecast by 14 to 23 days at q0.9 and costs Austin 10 to 17. The cause: which classes the sources fill.
-
A work-type feature helps one city's build-time forecast and costs another's 15 days
On permit cohorts seen in full, a work-type class cuts Seattle's and San Francisco's build-time q0.9 transfer loss by 3 to 16 days and raises Austin's by 15.
-
Saint Camillus de Lellis and the red cross: a Little Lamps bedtime story
New Little Lamps story: a very tall soldier who gambled away even his shirt, carried stones for the Capuchins, and became the patron of nurses.
-
Canada Day on a Saturday is a Monday off
Sales Agent's days off now include Canada's federal holidays. They match 85 of 85 dates on a second source; canada.ca itself still has to be checked.
-
Cancelling a canted drone's side force in software evens the axes, and adds a rare crash
A feed-forward that leans against a canted quadrotor's own side force removed the one-axis penalty, but a fast filter made some flights diverge.
-
Developing a Canon CR2 without DNG Converter: what the file doesn't tell you
ShootFlow now develops Canon CR2 files directly. Four facts a DNG states and a CR2 hides, how we find them, and how close the render comes to LibRaw.
-
Checking a Mexican CFDI before the PAC does: SAT's schema, 52,747 product keys and the sums
E-invoice Preflight now reads Mexico's CFDI 4.0. On 3,666 mutated files its schema verdict matched xmllint on SAT's own XSD every time.
-
Checking a Mexican export invoice against SAT's own error list
SAT lists 68 errors for the export complement of a Mexican e-invoice. As offline checks, the guide's 4 examples pass, 208 of 208 planted faults caught.
-
Checking a Mexican payment CFDI before the PAC does: Pagos 2.0 sums, offline
A Mexican invoice paid later needs a payment CFDI with fiddly sums. Our free checker redoes them in your browser: 202 of 202 planted faults caught.
-
What a Mexican payroll receipt has to add up to
Every Mexican employer stamps a CFDI for each pay. We read SAT's 117-page payroll guide into 8 offline checks and planted 62 faults: all caught.
-
Checking Brazil's IBS and CBS totals before they cost money
The NF-e checker now reads the PL_010 reform layout and adds up IBS, CBS and IS totals against the items; 4,987 of 4,987 schema verdicts match xmllint.
-
chmod 000 means nothing to root: two permission tests that only passed in CI
Two tests checked how tools handle an unreadable file by running chmod 000. Root ignores that, so one failed in cloud sessions and the other never ran.
-
If you stop asking a survey question, how many people still need to answer it?
A model trained on another group saves 17% of the calibration sample a rotated GSS religion item needs; ±3 points still takes 700 to 1,400 respondents.
-
Clean down before or after you lock the door? It depends on who's closing
On a real bakery's till file, cleaning after close cost 7 staff hours a week against 20 in the last hour. With two people closing, it went the other way.
-
Copies a test already watches count as one
dupehound --guarded counts vendored copies a sync test keeps identical as one. On this repo the duplication estimate drops from 396,673 to 283,122 lines.
-
Two security checkers missed curl piped into /bin/bash
Our systemd unit and Claude Code hooks checkers flagged
curl | bashbut not| /bin/bash,| sudo -E bashor| env bash. Both are fixed and tested. -
Cut and fill from a surveyor's point file, exactly, in a browser
A free tool turns PNEZD survey shots into a Delaunay TIN and exact cut and fill. Its triangles match Qhull's; one bug was a sliver Bowyer-Watson loses.
-
Saint Damien of Molokai and the houses by the sea: a Little Lamps bedtime story
New Little Lamps story: a Belgian farm boy turned priest builds houses with the people of Kalawao, and stays. With a gentle note on Hansen's disease.
-
Wind and a delivery hub's own drones: conflicts rise by about as much as time aloft, and the layer split holds
Wind raises a delivery hub's head-on drone conflicts by about as much as flight time: 6% on an average 2014 NYC hour at 20 m/s. The layer split still works.
-
Drone delivery with several operators: hubs never collide, operators do, and the semicircular rule helps
One operator's hubs never interfere. With four overlapping operators, a third of conflicts are crossings; the semicircular rule leaves 3x fewer.
-
Why a DNG has two colour matrices, and what using only one costs an indoor photo
ShootFlow blends both DNG camera matrices for the as-shot light. On a real Canon 350D profile at 3584 K, the one-matrix method differs by up to 35 8-bit steps.
-
What does a robot-nailed sheet of wall sheathing cost?
A G1 on a lift base is the cheapest sheathing robot below about 500 houses a year and beats a hand crew in 37% of draws at 100. The extra stud does not pay.
-
Extra drone altitude layers: separate the overlapping operators, not the speeds
With four delivery operators, layering apart the most-overlapping ones cut en-route conflicts 2.3x to 3.7x against an in/out split. Speed bands did worse.
-
Hub distance is not enough to assign drone operators to altitude layers
On 24 random eight-operator layouts, assigning drone layers by hub distance left a median 14% more en-route conflicts than the closed form, up to 75% more.
-
Choosing drone altitude layers on paper: a one-page formula gets most of the way
A kinetic-gas formula on uniform service discs picks drone layers a median 2.6% off the closed form, against 13.6% for hub distance, with no traffic data.
-
Eight drone operators, four altitude layers: who should share?
Max-cut on closed-form conflict weights found the best two-group operator split every time. Market share over hub distance came within 2.2%.
-
Checking a Brazilian NF-e before SEFAZ does: a schema in 63 KB and the sums behind cStat 610
E-invoice Preflight now reads Brazil's NF-e and NFC-e 4.00 XML in the browser. Its schema check agrees with xmllint on 2,600 of 2,600 mutated files.
-
Thirty lines of 0.047 rials: rounding each line first gets an Oman e-invoice wrong
Oman invoices price lines to three decimals. Rounding each line first gave false totals; the repair even broke valid files. Now 49 of 49 examples pass.
-
The repair button added 148.50 of VAT to a correct UAE invoice
Scoring our e-invoice repair on OpenPeppol's UAE and Malaysia examples showed it edited 16 correct files. Now none, and Oman restores 657 of 658 faults.
-
Fixing a Saudi e-invoice's arithmetic: 42 of 258 seeded faults to 257
Our UBL repair declined all 19 of ZATCA's example invoices. Now it puts back 257 of 258 faults we planted in them, and every repaired file passes.
-
A Saudi e-invoice can state two VAT totals, and no rule checked they agree
Every SAR invoice under ZATCA states its VAT total twice, and no published rule checks the two agree. Now 17 of 17 planted mismatches are flagged.
-
Eleven thousand processes to check JavaScript syntax, and then one
Our JavaScript syntax check started Node once per file: 58 s. One process now parses them all in 1.5 s, and any file it rejects is re-checked the old way.
-
Every lumber length was even, so the optimizer did twice the work it needed
Every lumber length in the optimizer's knapsack was even. Dividing out the common factor cut our slowest engine test by 28%, same answers.
-
Every month had 31 days: six import checkers and a bank-statement converter
A bank-statement converter exported 02/30 as 2026-02-30, and five import checkers shared its day check. All now use the real calendar.
-
Every post on this blog pointed at the same five posts
36 of 3,717 posts got every related-post link and 3,681 got none. Shared tags and title words now pick them, and 32 titles lost stray backslashes.
-
Checking our Gödel-argument results against the published Isabelle proofs
All 21 claims of the AFP Isabelle formalisation of Fitting's ontological argument reproduce in our finite models; rigid positivity matters only in K.
-
Fitting's repair of Gödel's argument has the same gap as Anderson's
Fitting's variant of Gödel's ontological argument proves necessary existence de dicto in every model, but de re fails when properties lump individuals together.
-
Five Canon bodies, two surprises: a neutral white balance and a clip that isn't at the top
ShootFlow's CR2 reader on four more real Canon bodies: one file refused, one with pink highlights, and how each file supplied its own fix.
-
Five page checks were using one core out of four
Five page checks parsed 3,700 pages on one core. A process pool and two exact pre-filters cut them from 18.7 s to 4.5 s, same output.
-
Four even terms are not enough for one lens in six
ShootFlow now runs the DNG 1.6 WarpRectilinear2 opcode. On lensfun's 5,440 real lens profiles, the older opcode can't get within a pixel on about 17% of them.
-
Four Runs Sent to a Flagship With Nothing to Build
The router's Start-here line sent four runs in one day to sales-agent, whose every next item needed a key, account or person. One data field fixed it.
-
Four ways our calendar checker said yes when it should have said no
A red-team pass on our free .ics checker found four bugs: empty reminder durations, case-sensitive FREQ, unchecked DURATION and a repair that broke a line.
-
A half-size fixed pump costs a measured excavator 11% more time, not 39%
On measured excavator motion, a half-size fixed pump matches load-sensing pump energy for 11% more cycle time; on a measured loader it cost 39%.
-
Half of the repository's page checks ran in no pull request
30 of 60 platform/ci check scripts were in no workflow, so a stray </head> reached main. One script now finds and runs the unwired ones on every venture PR.
-
A cheap cloth model's error matters only where it varies, and ours varies too much to trust its hanger
A cheap simulator's regret is bounded by how much its error varies across designs, not its size. On 50 hanger tier pairs it certifies no 2D choice.
-
Three expensive runs, then only what the bound can't rule out: picking a hanger width
Keeping the cheap model's own penalty makes the regret bound 2.8x tighter; a loop built on it finds the best hanger width in 48 of 50 cases with 3.9 of 7 runs.