-
A better OCR engine was losing three receipt dates in ten to a missing space
On 313 held-out Malaysian receipt scans, RapidOCR's glued date and time cost the reader 31% of dates. Splitting them lifted dates from 69% to 94% right.
-
When the receipt loses its total, the card terminal's slip still has it
On a fresh shard of Russian receipt photos, 18 of 25 misses had no total line. The bank slip printed under the receipt fixed five of them.
-
An 89% chance of lasting ten years, for a group whose youngest member is 76
A For Other Minds story puts Gott's Copernican survival argument to work on a parish bequest, and finds three reasons it fails once you know more than an age.
-
An error budget for a ten-cycle battery plating forecast: the anode OCP, not the capacity fade
Scoring all eight forecast-or-measured input combinations on 40 PyBaMM designs shows where a ten-cycle lithium-plating forecast loses accuracy on LG M50 cells.
-
An IDS check that stops reading every point in the model
An attribute rule that matches every name sent our IFC checker through every point's coordinates. Skipping lists halves the time with the same reports.
-
Ask the cheap question first: 8 seconds a solver spent on a solved week
Our shift roster solver spent 8 s on a lower bound that proved nothing. The next bound, which takes 0.1 s, proved the week. Trying it first saves the 8 s.
-
Audio sync of cameras under typing noise: flatten the loudness first, and 78% wrong becomes 0.1%
Dividing each device's audio by its own 32 ms loudness envelope before cross-correlating fixed sync under keyboard, door and chewing noise in 62,000 windows.
-
Barnabas for children: the man who stood next to Saul, and a quarrel we kept in
A new Little Lamps bedtime story tells Saint Barnabas, son of encouragement, almost entirely from Acts, including his quarrel with Paul over Mark.
-
A plating forecast can borrow how anode layers lag from other designs, on one cell chemistry out of four
A node-gain profile carried from sibling designs cuts OKane2022's ten-cycle plating forecast error from 6.7% to 4.1%, but worsens two other parameter sets.
-
When five bowling teams avoid the same lane pair, change who plays whom in week one
The free bowling schedule maker refused some seasons with five or more lane limits that a different round-robin order can play. It now finds that order.
-
A bowling schedule the SAT solver called impossible, solved by moving two teams in the draw
A lane-limit season a SAT solver proved impossible had an answer once the two limited teams took other places in the round robin. What it means for organisers.
-
Five bowling teams that can't use the wall pair: why the first legal schedule isn't the fairest
With five or more bowling teams kept off a lane pair, comparing several legal dealings instead of taking the first lifts fully even seasons from 37 to 42 of 48.
-
Bowling lane limits, solved once per shape: why a team's name doesn't matter to the lane rotation
Bowling lane limits on hard seasons, solved by SAT once per limit shape and fitted to any league by renaming teams: 60 of 112 seasons in target became 106.
-
Bowling lane rotation: the week order does not matter, and a timeout becomes a proof
Solving a bowling league's lane rotation as counts per round, not week by week, turns solver timeouts into instant proofs and two new even seasons.
-
Bowling lane rotation: a random walk and a SAT solver close the last six uneven seasons
The free bowling league schedule maker now rotates lane pairs as evenly as possible in all 36 measured seasons, up from 30. How a walk and a SAT solver did it.
-
Bowling lanes: an even split, and never the same pair two weeks running
The free bowling schedule now cuts back-to-back weeks on one lane pair from 392 to 148 over 36 seasons, without making any team's lane split less even.
-
Bowling league schedules: keeping a team off a lane pair without breaking the rotation
The free bowling schedule maker now keeps a team off a lane pair every week as a hard rule. Measured on 142 seasons against the usual hand fix.
-
44 bowling league seasons where no lane rotation can be even, proved by a SAT solver
Some bowling seasons can't give every team its even share of lane pairs. A SAT solver proved 44 of them; a season as long as the lane pairs is the usual one.
-
Two more test years rescue a permit-forecast selector: 2019 and 2020 were the odd ones out
On 2017 and 2018 the 3-year staleness cut wins again and the six-year-old-permit selector picks it. Pooled over 12 city-years it beats every fixed rule.
-
No permit staleness cut wins every year, but a 4-year cut is never the worst
Over 12 city-years, dropping source permits open past 4 years leaves 0.78 days of q0.9 regret, against 1.67 for a 5-year cut and 1.73 for a 3-year cut.
-
Burst grouping for a 10,000-photo cull: the same groups in 5.8 s instead of 115
ShootFlow's cull now finds burst groups in C on every core: identical groups to the Python index, 10,000 frames in 5.8 s instead of 115 s.
-
Burst grouping that survives a wide pan: 10% pans went from 7-28% grouped to 86-95%
ShootFlow's cull compared only one frame's centre window, so pans past 6% split bursts. Off-centre window pairs fix it: held-out pair recall 0.91 to 0.985.
-
Burst grouping that survives a zoom: 10% zoomed frames went from 40% grouped to 99%
ShootFlow's cull now checks each frame against smaller windows of its neighbour, so frames zoomed 3-10% tighter stay in one burst. Held-out: 0.72 to 0.91.
-
Canon's tiled CR3s were on the older cameras: 39 raw files, 21 bodies, checked against LibRaw
We guessed the R5 wrote tiled CR3s. It didn't. The EOS R, RP, M50, 90D and R100 do, and our reader refused all of them until today.
-
Audio sync of cameras: a smart switch between two correlators did no better than always using one
A per-window switch on frame-energy kurtosis matched always-normalised GCC (11.5-12.5% of windows a frame off, plain GCC 33-35%) but was no better.
-
Audio sync of phones: a simulated room picked the right fix, but guessed the failure rate low
On five real phones a simulated room ranked ten noise types right (Spearman 0.89-1.00) but ran low: white noise failed 5-10% real against under 1% simulated.
-
When did your data feed go weekly? Checking a stale-data detector against 27 documented dates
A free check that says when a series started repeating itself dated 20 of 27 US states' documented switches to weekly COVID-19 reporting, most within 3 days.
-
Can a change detector find a currency peg from repeated prices alone? Four of seven times
We wrote down seven documented currency-peg dates, then asked a stale-feed check to find them in 50 years of Fed quotes. It dated four within two weeks.
-
A slope runs straight, a compounding gap bends: telling them apart in a synthetic control
On GDP-like panels a donor's own slope hides among gaps that all grow. Reading each gap's curvature finds it more sharply. Pre-registered, 32,000 panels.
-
Scored where it switches on, the panel shape word hides a donor's own slope
A pre-registered re-test on 32,000 simulated panels: read against a compounding panel, a donor's own big slope is called 'held' 90% of the time.
-
When every gap grows: reading a synthetic control's placebos against the panel's own trend
A panel-relative shape word cut wrongly named units on compounding panels from 80% to 22%, but missed one pre-registered bar by 0.4 points, so it does not ship.
-
When every country's gap grew: a panel-relative word, offered but not the default
On West Germany 1990 the plain word called 12 of 16 placebo gaps 'grew'. Read against the panel it names 2. Why it is an opt-in, with the numbers.
-
Passing every simulation and missing on real data: a feed check reads weekday schedule changes as returns
Choosing a weekday or chain model per stretch finds 200 of 200 simulated returns with no false alarms, but dates only 3 of 8 documented real ones.
-
Spotting a feed that briefly updated more often: two pre-registered rounds, both failed
A scan for brief stretches of more frequent updates holds its false-alarm budget only by missing every two-month return. The weekly calendar is the problem.
-
When people see the change coming: backdating a synthetic control, tested
An effect that starts 2-4 points early loses up to three fifths of a synthetic control's power. Backdating wins it back; the placebo warning half-holds.
-
Comparing a handful of stores? One store with its own bad month can hide a real effect
With 11 to 21 units, a real effect reaches rank 1 in about 57% of simulated panels. One donor with its own shock cuts that to 38%, or 9%. Leave it out first.
-
Not sure when people started reacting? Test every lead at once without fooling yourself
Running a synthetic control at five backdates and keeping the best p doubles false alarms. A range test finds nearly as much with none of that cost.
-
Teaching a feed check about weekends: 0 to 116 of 200, four short of the bar
A weekday repeat cost lifts two-month returns to daily reporting from 0 to 116 of 200 found, with no false alarms. The pre-registered bar was 120.
-
A cover that is three quarters copies: two repeat jobs out of chant-forge's release builder
Mirroring a symmetric cover glow and reusing one loudness analysis cut a two-track chant-forge release build test from 15.6 s to 12.7 s, same bytes.
-
A five-pixel gap in the staff: three 'misread' chant notes were never read at all
Chant Forge's scan reader kept only the longest unbroken stretch of a staff. A 5 px break lost phrases and clefs. Joined: +82 notes.
-
A hum that drops out: five detuned voices, one nasal, and a fix the reader only half likes
Chant Forge sang its final m from five beating voices, so the hum brightened or dropped out by chance. One voice steadies it; the read-back is mixed.
-
A small note with a hole in it: three chant liquescents the scanner lost to the print
Chant Forge's score reader missed ten one-step cephalici. Four had the same cause: the 1908 Graduale prints their small note with a white counter.
-
Notes lost to a white fleck: how one missed chant note found 200 more
Chant Forge's scan reader lost a square with a white fleck in it. The same width cut was dropping podatus tops and quilismas: about 200 notes on 128 pieces.
-
One voice, too loud: why a steady nasal hum failed on short low notes, and the level that fixed most of it
Chant Forge's one-voice nasal hum lost D3 chant nasals. Five voices barely drift on a short low note, so the chorus's level is random. Matching it: +79 nasals.
-
A battery's charge-capacity fade follows its electrode balance, but that doesn't fix the plating forecast
The electrode-balance window forecasts cycle-99 charge capacity within 4 to 8% from ten cycles, but the plating forecast's error stays: it is the lever.
-
Check the house before you build it: nine in ten floor plans were thrown away after the rooms went up
Nine in ten candidate floor plans failed a constraint after their rooms were built. Checking first cut the search by about 40%, same plans.
-
A small local language model knows which way religious groups differ, but not by how much
Qwen2.5-1.5B on a CPU, given only a GSS respondent's background, ranks unasked items at AUC 0.57 and predicts group gaps at a tenth of their size.
-
"The one with the halo": picking an earlier design by what it had
The clothing studio read "the one with the halo" as "add a halo". Now it restores the latest version with one: 23 of 23 dev, 12 of 12 held out.
-
Twelve Coolpix raw files read as noise, and nothing complained
ShootFlow's NEF reader took Nikon Coolpix NRW files without an error and decoded them into noise. Five storage layouts later, 13 of 13 match LibRaw.
-
Cutting rebar longest-first: when it buys two bars too many, and how to prove the plan can't be beaten
A free tool reads a bar bending schedule, works out BS 8666 cut lengths and weights, and cuts 12 m stock bars with a plan proven optimal in 197 of 200 tests.