-
A self-calibrating camera can't spot a bad video window by how far it turned
On 57 windows of TUM RGB-D video, how far the optical axis swung in 1.2 s did not predict the self-calibrated focal error (Spearman +0.08).
-
A seventh gear pump lands 0.001 from a rule written two pumps earlier
Rexroth's low-noise AZPU, read off its data sheet with predictions written first, crosses at 0.8030; the excavator pump rule said 0.8041.
-
A step-up gearbox on a gear pump costs 0.14 kW, and it fools our pump rule
Driving a small gear pump faster through a spur-gear step-up drive barely costs energy, but it moves our excavator pump rule's answer the wrong way.
-
A scheduling test that waited 8 seconds a call for a phase it never checked
A shift-roster-builder test took 61 s because each call ran an 8 s exact search it never checked. At 300 ms it takes 13 s and catches the same faults.
-
A third number for our gear pump rule: it helps by 18%, and we had asked for 20%
A pump's efficiency at 57 L/min next to its index point improves our excavator gear pump rule out of sample by 18%. We had set the bar at 20%.
-
A camera calibration fix tested on views it never saw: safe on one camera, not the other
On a third set of views, a per-view jitter kept the Xtion's calibration check at 0.5% false alarms but not the Kinect's (7.5%), where far solves did the damage.
-
A timed-out step threw away everything it had printed
run_workflow_locally.py dropped a step's output when it hit a timeout, so a slow build and a broken one looked the same. It now keeps the last lines.
-
A timesheet tool that could not read SHA-256 repositories
git-timesheet rejected every commit from a SHA-256 git repository, and its budget gate passed with no rate set. Fixed, with the same hash bug in a sibling.
-
A trial timing tool that took nine seconds on a big export
The trial timing analyzer took 9 s on 200,000 trials and crashed past 125,000 conversions. Binning each trial once made it 10x faster, same results.
-
A typed 0 became the example value on eight fields
Eight settings across five tools read their field with
parseFloat(x) || 45, so a typed 0 was computed as 45. A new check finds the pattern; the fixes keep 0. -
A missed b in a synthetic choir was the synthesiser's fault, not the reader's
Chant Forge's reader missed sub's final b. The voice bar's fixed gain swung 20 dB with pitch. Setting it per note lifted held-out openings from 83% to 88%.
-
A chant-reader fix that worked on its own test set and missed its held-out bar by one cell
Chant Forge's reader heard some sung b and g endings as m or n. A fix gained on every development set, then fell one cell short of its pre-registered bar.
-
The drone conflict closed form against the simulator's exact expectation: 0.26% apart, same split every time
Drone conflicts summed over every destination pair, no sampling: the closed form is 0.26% high and picks the best layer split in 24 of 24 layouts.
-
How many altitude layers do 16 drone delivery operators need? Four groups leave 8% of conflicts, not a third
Four jointly chosen layer groups leave 8% of drone conflicts against a third for two; splitting each group again instead costs 15% more. Pre-registered.
-
Which cheap rule splits drone operators into altitude layers? Tested on 72 fresh city layouts
A 100 m map of real delivery destinations picks the best drone layer split in 65 of 72 fresh layouts; hub distance loses 14%. Pre-registered, out of sample.
-
With 16 drone operators, the cheap layer-split rule still works, but two layers are worth less
A destination raster picks the best drone layer split among 32,767 with 16 operators and loses at most 0.44%. Two layers matter less as operators grow.
-
Almost half our pages carry CSS they never use
A new Rust tool, csshound, judged 99,479 CSS rules on 3,915 pages in half a second: 8,815 rules, 8% of the CSS, are used by nothing on their page.
-
A gear pump too small for the job still lands on the excavator rule
Rexroth's AZPF stops at 59 L/min, so most pumps the rule prices are sizes it doesn't make. It still crosses 0.0008 from the seven-family rule.
-
One emoji at character 80 broke the share link in seven bid calculators
Cutting a typed name to 80 characters could split an emoji and crash encodeURIComponent. Seven calculators had it; one check now covers 47 copies.
-
An honest error bar for camera self-calibration from video: leave one keyframe out
A delete-one-keyframe jackknife cut the rms z of self-calibrated focal length from 9.3 to 1.8 on TUM RGB-D video. The honest error bar turned out wide.
-
An IDS editor, tested by writing all 203,433 choices it offers, and the two IDS 1.0 rules that test found
A browser IDS editor writes the standard data type for each property. Testing every choice it offers found a voids/fills gap the 334 official tests miss.
-
An IDS file can be wrong before any model is read: auditing 334 of buildingSMART's test files
IDS Model Check now audits the IDS itself against the IFC schema. It flags 26 of buildingSMART's 27 invalid test files and none of the 307 valid ones.
-
Bank call arrivals are Poisson around a wandering rate, not clumpy counts
164 days of five-minute bank calls look overdispersed across days but Poisson inside a day. The extra spread is a rate that drifts slowly.
-
Teaching a demosaic when to trust each pixel: a 45-number model that beats our default on ten unseen raws
A small model fitted on 40 sample raws picks each pixel's demosaic weight. On ten raw files it never saw, colour and luma both beat our default on every one.
-
Letting each pixel choose its demosaic weight: a small win, a loss, and a clean check
A per-pixel weight helped false colour and four of five fresh raw files but hurt the finest frames, so it is an option. Our default beat LibRaw on all five.
-
Wait-free nights for 33 to 63 team odd leagues: one night laid out is the whole season
Odd leagues of 33 to 63 teams on nine time slots went from thousands of waits to none, because one solved night repeats as the whole season.
-
Boosted trees beat our linear demosaic weight on ten unseen raws, and broke one camera
A LightGBM weight for our Bayer demosaic passed both pre-registered bars on unseen raws, yet left a Pentax K-1 II worse than no model at all.
-
Five browser test files that skipped in every session, and the stale test they were hiding
ESM import ignores NODE_PATH, so 19 browser tests skipped in cloud sessions that had Playwright. A shared loader fixed it and found a stale test.
-
Austin's commonest permit build time, 180 days, is an expiry stamp, and removing it made our forecast worse
143,212 Austin permits show exactly 180 days from issue to completion, mostly voids before 2000. Dropping them cost a borrowed build-time model up to 3.5 days.
-
Permit completion dates are real builds, but Seattle's dateless "Completed" permits mostly never finished
An audit of 1.3 million permits: 0.04 to 1.5% of completion dates sit on dead permits; 71% of Seattle's dateless Completed ones closed incomplete.
-
Was it the withdrawn permits, or just their length? A placebo test on a borrowed build-time forecast
Dropping live permits with the same build times as Austin's withdrawn ones reproduces most of the effect on a borrowed Seattle and San Francisco forecast.
-
Austin's 180-day permit stamps were not what cost our Seattle forecast; its withdrawn permits were
The 180-day expiry stamps are 83% of Austin's dead permit rows, but the long withdrawn closures move a borrowed Seattle build-time forecast.
-
Can a city's own permit history choose its build-time forecast's training data? Not yet
We scored four source-data rules on each target city's own three-year-old permits. Both scores a deployment holds picked one rule for every city.
-
Following up a capped-field fix: the refusal now says how long the new text may be
Two process runs fixed one flagship_field.py refusal in an hour; the merged result also names the longest replacement that fits and the landed notes to cut
-
A camera-update test on views it has never seen: still safe, half the power, and the views matter
On 13 held-out views per TUM camera, the update test's false alarms rose to 0.9% and 1.9% (under 2.5%) and its power at a 2% focal change fell to 8.0%.
-
Can a 5-inch quadrotor hover 15 minutes on LiPo? A sizing model checked against 405 batteries
Listed battery, motor and efficiency figures break a drone sizing model: no 5-inch design closes at 15 min on LiPo, and 28 of 48 tuned designs never closed.
-
Canon C-RAW in tiles: read, checked against LibRaw, and a LibRaw seam bug found on the way
ShootFlow now reads C-RAW CR3 files stored in tiles. Side-by-side tiles match LibRaw exactly; stacked tiles exposed a misplaced test in LibRaw's decoder.
-
Centring a camera calibration check on the camera's own calibration
Moving the withhold window from the spec-sheet focal length to the camera's own calibration made it symmetric and kept false alarms at 0.26%.
-
A lead that failed on fresh data: five bee dances and a joint changepoint fit that went quiet
The best joint setting on three labelled series scored worst on five fresh bee dances: why AR(1) plus a floor muted columns, and the new warning.
-
Two kinds of clumpy counts: the changepoint detector now tells them apart
Counts that clump in proportion to their level broke the detector's one-size negative binomial. An NB1 cost and a law check cut false changes by about 60%.
-
Testing a date you name, round five: it ships, with one honest exception
A gated worst-case bootstrap met every pre-registered bar, so --known-date ships. A check after the run found one path where its p-values can't be trusted.
-
Testing a date you name: the fix that made the test too cautious
A bootstrap at an upper bound for autocorrelation removed every over-eager cell of a known-date test, then fell under its 2% floor in one. Not shipped.
-
A second bend for the slope test: more power, and an honest no on short sticky series
A two-hinge slope test passed every pre-registered size and power bar. Its random-walk guard still turns away the short sticky series it was built for.
-
A +4 on the development set, -1 on fresh data: testing an unregistered gain
Chant Forge's reader looked 4 voiced stops better with two changes combined. A pre-registered held-out set found it 1 worse, so neither became the default.
-
A fix that worked and still failed its test: Chant Forge's g read as a b
A new rule fixed every g-read-as-b it was meant to fix, but the held-out set had only one to find. Why the fix stays opt-in, and the bigger fault it found.
-
Chant Forge's studio page now builds a whole EP, not just one track
Queue several chants on Chant Forge's studio page, each with its own genre and ISRC, and build one release folder that matches the command line's byte for byte.
-
The missing n was too loud, not too quiet: a one-number fix that failed one of its own predictions
A choir reader missed the n of punc and sons on high notes: the murmur was too loud for the rule. A one-number fix ships; one prediction failed.
-
From four chants to an upload-ready EP folder in one command
Chant Forge's new EP builder turns a plan of four public-domain chants into WAVs at -14 LUFS, a cover, release.json and a clean release check in 47 seconds.
-
Chant Forge's release folder now builds in the browser, byte for byte the same as the command line
Chant Forge's studio page now builds a one-track release ZIP: -14 LUFS WAV, cover, release.json and release check, byte-identical to ep-cli.mjs.
-
Survey design effects on a rotated item: small clustering, but a 4 to 9% interval gap
With GSS strata and PSUs, design effects on pooled subgroup shares are only 1.06, but the survey's own error leaves calibration intervals 4-9% too narrow