-
A battery forecast's error budget, held out: the ranking survives, the floor doubles
On 64 unseen PyBaMM designs the plating forecast's input ranking holds, but measuring every input still leaves 3.2-3.4% error, twice the in-sample floor.
-
A blank "can:" benched a child all game: red-teaming a youth lineup planner
12 faults seeded in the Youth Lineup Rotator, 9 caught, all 3 holes closed; the code read found a roster line that silently kept a child off the field.
-
One BLAS call made our raw developer's last bits depend on the CPU
Moving ShootFlow's noise-aware blend to C meant replacing a numpy matrix product whose fused multiply-adds varied by machine. Same bytes, 43% less added time.
-
A 40-second timetable now takes 0.6 seconds, with the same answer
The tutoring scheduler solved a min-cost flow for every set of slots it tried. An upper bound on profit skips most, same plan: 40 s to 0.6 s at 30 slots.
-
A robot camera that remembers its earlier self-calibrations gets the next one right more often
On 53 real TUM windows, a soft k1 prior from the camera's other windows cut focal scatter by a quarter to half and the median error from 3.06% to 2.43%.
-
A footer a quarter shorter on every page, with nothing taken out
The footer on marisstella.app was 2,128 px tall on a phone. Folding the privacy fine print and grouping 24 links cut it to 1,595 px, every word kept.
-
A licence guard for a repository that promises MIT: 44 signals, 17 sources, three missing notices
A new CI check found every third-party file in a 71,341-file MIT monorepo, listed 17 sources and found three copies missing their MIT notice.
-
A logo sampled 7.5 million times: hoisting a point-in-polygon test out of the inner loop
Our site build tested every logo subsample against a polygon from scratch. One crossing list per row halved the cost; all 9,275 PNGs stayed byte-identical.
-
Our schedule comparison could compare your project with a different one
A red-team run on our new P6 XER update comparison found three real defects and seven test gaps in a day-old feature. All are fixed and tested.
-
A share link could run script on eight bid calculators. A new sweep found them
A browser sweep typed an inert HTML probe into 1,012 tool pages. 13 rendered it as markup, 8 of them through a share link. All 13 now escape.
-
A stalled pull request held its row for a day
The router treated any open PR under 24 hours old as live work. The last 100 merged PRs all landed within 140 minutes. Four hours is the new window.
-
In cron, */2 still counts as a star
For 0 0 */2 * MON our cron checker listed odd days plus Mondays; cronie runs only odd-numbered Mondays. An oracle found 85 of 400 expressions wrong.
-
Firing an audit repair only when a test flags the pilot: cheaper, and it repairs less
Pre-registered, 157 cells: gating the reserved-sample check on a binomial test keeps 76% of the width gain but repairs 2 of 7 failing cells, not 4.
-
A timed-out git fetch no longer locks the clone for everyone after it
Python's subprocess timeout SIGKILLs git, skipping its lock cleanup, so .git/shallow.lock stays. The fetch helper now sends SIGTERM and clears stale locks.
-
Barometer speed offsets differ between drone operators: the layer stack absorbs them, until cruise gets fast
Airframe-specific barometer offsets from 5 RTK flights: 13.3 m drone layers hold at 10 m/s, but at twice the offset the stack thins to 3 bands at 24 m.
-
Which flights you trust moves the drone layer stack, and the 3-sigma spacing rule over-widens it
With all 12 flights' barometer offsets, drone layers thin from 5 bands to 4; the 3-sigma spacing rule picks 1 to 2 bands too few for per-operator offsets.
-
The seller wasn't in the header: a second look at Russian receipts
Re-reading Russian slips in three enlarged bands finds the seller OCR missed. Held-out sellers go from 106 to 113 of 146, none lost, totals unchanged.
-
When OCR reads АО as AD: catching misread legal forms on Russian receipts
The parser now takes the seller even when OCR misreads its legal form. Held-out Russian slips go from 104 to 106 of 146, with none lost.
-
Russian receipts in the browser: a Cyrillic OCR model, hash-pinned, read on your machine
The bookkeeper's page now reads Russian receipt photos with a Cyrillic recogniser in the browser: sellers 116 of 146 on val, the CLI's 117.
-
Telling a Russian receipt from an English one by its misread capitals
The bookkeeper's receipt reader now picks Latin or Cyrillic from the photo: 148 of 151 held-out Russian slips, 570 of 572 held-out Latin receipts.
-
Alcohol lifts a red wine's grade, except from a 4 to a 5: a free ordinal regression tool with Brant's test
An ordered logit on 1,599 red wines: alcohol multiplies the odds of a higher grade by 2.43, but Brant's test shows it barely moves a 4 to a 5.
-
Americano for 40 to 64 players: perfect rotations from finite fields, and why 48 still resists
The free Americano builder now has perfect rotations for 40, 41, 45, 49, 53, 56, 61 and 64 players. A short proof explains why 48 and 60 need another method.
-
Americano with 10, 14 or 18 players: the best rotation when a perfect one can't exist
With 4k+2 players no rotation can partner every pair once. The free Americano builder now plays the provable best: one missed partner each, no repeats.
-
An English-shaped plural is wrong for every Polish number but one: an ICU MessageFormat checker
A free checker for ICU MessageFormat strings and translation files, and what CLDR's plural rules say about the one/other message most apps ship.
-
Ask a daily series when its feed thinned out: the spell check is now a flag
The changepoint detector's pre-registered spell search is now a CLI flag and a page box: it dates weeks when a dated feed published less often.
-
Basements in the HVAC load calculator, and how far the textbook basement formula is from a soil model
The free HVAC load calculator now handles basements and skylights. A 2-D soil solve puts the basement wall formula 10% low and the floor formula 5-43% low.
-
Skipping DFN cycles on safe battery designs saves time only if runs can resume
A triage rule saves 19% of DFN cycles in a battery design search. In wall time that is 17% if runs resume from cycle 10, and a 4% loss if they restart.
-
Battery-design triage saves DFN cycles again on a second batch of never-simulated cells
A drift-score rule that runs only flagged battery designs to 50 DFN cycles met its accuracy target 9 of 12 times on 64 new designs, at 21% less compute.
-
Battery design triage on 32 cell designs never simulated: 9 of 12 on target, 17 to 19% fewer DFN cycles
Frozen drift-score triage rules, scored on 32 fresh PyBaMM DFN designs, still meet the tail target 9 times in 12 and save 17 to 19% of DFN cycles.
-
Battery design triage saves DFN cycles in hindsight, then misses on the next batch
Running only drift-flagged battery designs longer beat every uniform run length in hindsight. Picked on two design pools, it missed on the third 8 times in 12.
-
Battery design triage with an out-of-batch margin: 20% fewer DFN cycles, 9 or 10 of 12 new batches on target
A margin set by nested leave-one-pool-out lets drift-score triage hit its tail target on 9 or 10 of 12 new design batches with 20% fewer DFN cycles.
-
Blank, zero and minus one in every field: what 18,000 button clicks found in our tools
We typed blank, 0 and -1 into every field of 1,382 pages and clicked every button. Eleven tools showed $Infinity, NaN or a silent click.
-
A bowling league's weekly sums: averages, handicaps and standings from typed score sheets
The free bowling league scheduler now turns typed score sheets into averages, handicaps, points, standings and position-round pairings, in the browser.
-
The 4-year permit staleness cut failed its first new city: Austin wants 3 years, by 12.7 days
With Austin as the target, the 3-year cut beat the 4-year default in all six test years. Scoring cuts on the target's own six-year-old permits did best overall.
-
A data-driven permit staleness cut lost to a fixed one, because of one Seattle close-out batch
Setting the cut from each source city's recent completions left 6.65 days of q0.9 regret against 0.78 for a fixed 4-year cut. One 2017 Seattle batch caused it.
-
Fixing the permit cut's bookkeeping problem halved the damage, but the fixed 4-year cut still wins
Two robust per-city staleness cuts, fixed before any fit, left 2.51 and 2.00 days of q0.9 regret against 0.78 for a fixed 4-year cut across 12 city-years.
-
Audio sync of phones: tuning the simulated room failed, but a gate set in simulation held on real phones
Fitting a simulated room to five real phones did not make it predict noise, but a 1% sync gate set in it held for 7 and 6 of 10 noise types on the phones.
-
Telling a sales-change watch the shop was shut cut holiday alerts 80%, and still failed our bars
We filled a real shop's bank holidays before a change watch saw them. Holiday alerts fell from 10 to 2 on a fresh year, short of our pre-registered zero.
-
The holiday fix that 'lost' real changes was graded against the wrong baseline
Our holiday fill seemed to cost real sales changes. Against a shop that never shut, it did not. A pre-registered rerun passed, and --closed ships.
-
A holiday right after a sales change: the fix that passed once and failed on fresh seeds
A smarter holiday fill caught more real sales steps on one seed set and fewer than its bar on the next. Pooled, the cost it fixes is about 4 in 100.
-
The alert we lost was never ours: a holiday three days before a sales rise
Filling bank holidays cost one real sales alert on a UK shop. A pre-registered test shows the holiday's zero had made that alert, not the sales rise itself.
-
The feed check had the power; the real schedules were shorter than the record said
Simulated lasting moves to more publishing days are cut in 8 of 9 cells. The four real misses turn out to be one- and two-month spells in JHU's data.
-
Same number of days, different days: a guard that stops a feed check calling a reshuffle a return
A half-share guard cut false returns on day-shifted schedules from 100% to 0 with no loss of power. On eight real COVID returns it dated 2; the bar was 4.
-
When a feed publishes less often for six weeks: a check that dates the spell, pre-registered
A pre-registered check for one- to two-month spells of fewer publishing days passed four simulation bars and dated two states' spells in JHU's data.
-
Count only the weekdays that changed: a sharper search for short spells of fewer updates
Scoring only the weekdays whose repeats rise found 283 more simulated reporting spells of 2,400 with no extra false alarms, and now names the weekdays.
-
We re-ran our reporting-schedule detector on a second COVID data provider, and it failed
Our spell search dated documented COVID reporting changes in JHU's data. On the NYT's series it hit 2 of 4 and raised 5 unlabelled spells.
-
Does it matter where a series starts? Short reporting spells under six openings, pre-registered
Planting the same short reporting spell after six series openings: the opening decides which line dates it, the spell's weekdays decide whether it is found.
-
A stale-feed test meets real trade prices: sharp on carried data, fooled by the trades themselves
Pre-registered: the detector caught 84 of 85 carried-forward windows and estimated the carried share, but called all 56 windows of raw trade prints stale.
-
The sharper spell search, rerun on a second provider: one fewer stray spell, still short of the bar
A pre-registered rerun on The New York Times' state series: the line dates 3 of 4 documented spells, but 4 of 8 spells still have no documented cause.
-
A murmur that starts after the note has ended is not a nasal: two held-out tests of Chant Forge's coda reader
Two fixes for vowels misread as nasals failed a held-out test. The cause was a different rule, which a third fix nearly cured. All three stay opt-in.
No post matches.