-
Three false errors on Canary Islands invoices, found in the AEAT's own rule text
Our VERI*FACTU checker rejected valid IGIC records three ways. The AEAT's validation document showed why; all three are fixed, and two new checks added.
-
Can a cheating training worker forge Adam's moments? A weight check is fooled, a moment check is not
Two zero-gradient forgeries hide skipped steps from a weight-only replay audit under Adam; checking the committed first moment catches them.
-
Can a cheating training worker save compute with sub-batch gradients? Under 0.4%, and it still gets caught
A worker who computes some steps' gradients on part of the batch saves at most 0.4% of its compute before a moment-checking replay audit catches it.
-
A confirmatory blind test of a cheap chip-flow ranking: 21 wins and 9 losses on 33 designs
Twelve new SKY130 designs, one clock each: the clipped ladder beats random routing on 8 and loses on 4. Pooled over 33 designs, p = 0.043.
-
Testing a fix that follows from the diagnosis: a delay check on the chip-flow ladder's top picks fails
Routing the ladder's top 3 and checking whether the smaller netlist is slower flags 47% of cells where the ladder is perfect, and 45% where it fails.
-
A cheap chip-flow ranking, tested blind on right and 256-bit shifters: 21 wins in 25, and the mean edge holds
On 25 new SKY130 designs the clipped ladder beats equal-cost random routing 21 to 4; over 58 blind designs its mean edge is -2.47 points, interval below zero.
-
Why a cheap chip-flow ranking fails: the area half is right, routing makes the small netlists slow
On 36 blind SKY130 cells, mapped area ranks routed area at tau 0.95, but where the ladder fails its picks route 3% to 27% slower than the optimum.
-
Village survey respondents forget about one tie in six, and the seeding advice still holds
Capture-recapture on converse survey questions puts tie recall at 84% in 43 villages; at that recall a half-village survey still seeds near a census.
-
Watching a slow launcher wheel past the verdict: 500-750 rpm drops from 1.84x off to under 1%
A tennis feeder's slowest wheels had three or four revolutions before the shot's verdict. Watching them a little longer brought every ball's drop within 1%.
-
When Elo said 58%, home teams won 56.5%: a free probability calibration check
A free no-upload calibration check for probability forecasts. On 2,930 NFL games, Elo passes the curve test yet overforecasts home wins, mostly in 2019-20.
-
When OCR misspells TOTAL: letting the receipt's own sum vote
A receipt photo read four times can lose the word TOTAL in every reading. Its subtotal plus tax, printed on the page, now gets a vote.
-
When OCR spells Subtotal "Suototal": six more receipt taxes from the receipt's own arithmetic
Misspelt subtotal labels and tax lines with lost decimals hid the tax on receipt photos. Checking each guess against the total read six more of 185.
-
Where the excavator fuel optimum picks its gear pump, and why reading the pump there still gives 18%
Over 2304 excavator geometries the best fixed gear pump sits at 58 to 90 L/min. Reading pump shape there matches our earlier 18% gain, not the 20% we asked for.
-
Which side should a corridor robot keep? Three crowd models agree only on the wrong answer
In three fitted crowd models, a robot that keeps the side opposite people's convention is worst on every measure; the models disagree on which policy is best.
-
Why a 29-team league can't be one night turned nine times
A short proof that no renaming of order k turns one night into a 3k+2 team season, plus three routes measured on 29 teams on 5 courts.
-
Why a slow tennis-launcher wheel never settled, and the gate that let balls go 13% off
A 510 rpm launcher wheel rang for ever because the tachometer's delay grows as the wheel slows. Scaled gains and a longer gate fixed it on the simulator.
-
Why our chant reader heard every sung /i/ as a z, and how we fixed it
Chant Forge's consonant reader heard /i/ as a z. The cause was five synthetic voices beating out of phase. Held-out /i/ reads went from 69% to 96%.
-
Why our chant reader heard a sung p as a k: one 8 ms noise draw is not a spectrum
Chant Forge's reader took sung p bursts for k: a single 8 ms noise burst has a random spectrum. One rule took a held-out part from 18 such misses to 0.
-
Why lithium plating slows down in LG M50 battery models: look at the particle surface, not the rest voltage
In PyBaMM's DFN model the LG M50 cells' plating slows 3.3x over 100 cycles because the surface OCP by the separator rises 38 mV, not the 0.5 mV seen at rest
-
Why the tennis feeder's spin reading was off by 8% when its speed reading was right
The open tennis feeder reads spin from wheel-speed drops. The 8% spin error came from leverage and a speed bias; faster spin shots halve it.
-
A wrist force sensor fixes most of a nail gun hose's pull on a robot, but not the last half millimetre
With a wrist F/T sensor, 0-20 N of hose drag costs a humanoid 1.6 points of clean sheets, not 92. Joints at the friction threshold still leave 0.3-0.8 mm.
-
03.06.2026 is June, not March. Red-teaming the delivery commission analyzer
Ad-only rows vanished from a delivery analyzer's totals, and 29 tools read 03.06.2026 as March. Both fixed; a CI check now tests every copy.
-
16-bit numbers brought 26-team leagues into the schedule page
Running the break tables as 16-bit integers made the league scheduler's break-floor search 1.6x faster with the same results, so 26-team leagues now get it.
-
The blog filed 415 bedtime stories under Tools
Saints' bedtime stories and science fiction showed a 'Tools and projects' badge on the blog. A Stories topic now holds 415 of them.
-
On 49 tool pages the header started at the very edge of a phone screen
On 49 of 1,351 tool pages the header logo and links touched the left edge of a phone screen. A padding rule meant for another layout was the cause.
-
79 pages told visitors they needed JavaScript to calculate. They had nothing to calculate.
The build's own embed box and search form put a false 'needs JavaScript' banner on 79 pages. Fixed, plus a RAW support page for ShootFlow.
-
A bedtime podcast that could not be heard: voicing Saints by Candlelight again
Saints by Candlelight had 126 episodes and audio for 19. Five more are now voiced offline with Kokoro, 7 to 11 minutes of compute each, free to every listener.
-
A bend cuts the stickiness; a random walk barely does
Pre-registered on fresh seeds: a guard that asks how much a bend cut the stickiness finds 83% of strong bends at 95 points (was 15%), with few false alarms.
-
A slope-change test found a bend in the Nile. It was a step.
Pre-registered: a step-versus-bend check got the Nile right and failed on synthetic bends; the slope test refused 7 of 8 short real series.
-
A click hidden in every final -m our synthetic choir sang
Teaching Chant Forge's acoustic reader to hear /m/ after /u/ turned up a 10 ms glitch in every final nasal its synthetic choir sang.
-
A crowd simulator that trusts the keep-right rule too much
Fitted to nine real corridor runs, a JuPedSim social-force twin kept 78% of walkers on their right once each had a side exit. Real people kept 54%.
-
A current cutoff saves the tennis lift's motor, not its flight
A shunt and comparator let the tennis feeder cut a jammed ball lift in 30-90 ms. The simulation also showed what no cutoff stops: the rotor's inertia.
-
A DONE line from the last phase: making a log wait tell phases apart
Two 10-minute waits returned at once because phase two appended to a log that already said DONE. The waiter now checks who still holds the file open.
-
A gear pump the rule never saw: Rexroth's AZPG needs the second number
A fifth gear-pump family, from Rexroth's AZPG data sheet, crosses at 0.804. Efficiency plus drag predicts that within 0.005; efficiency alone misses by 0.017.
-
A final l is not a hum: teaching a consonant reader to look at the coda itself
Chant Forge's acoustic checker heard every final al as a nasal. Comparing the coda with itself, not its vowel, fixed it; the same rule failed on openings.
-
Agents guessed compute. Now the tools server hands them a call that ran
describe_tool now returns a call it has just run on each tool's own sample inputs, so an agent's first call works: 769 of 1,350 tools get one.
-
A ball-lift flight with a pad that lifts off the belt round each pulley
Why a welded cleat alone nearly doubles a ball lift's belt pull, and the printed cap with a loose pad that fixes it without stopping the belt wrapping.
-
A fourth crowd model, from another code base, keeps right too
Fitted to real corridor runs, the anticipatory power law kept 97% of walkers on their right under assigned exits. People kept 54%. A yielding bias: 70%.
-
A Hungarian cancellation has to add up to zero: checking NAV corrections against their invoice
NAV checks a correction or cancellation against the invoice it refers to. Drop both into the checker and it now runs those checks offline.
-
A jaw coupling's free play rang the tennis lift's start. A 100 ms ramp fixed it
Modelled as a spring with play, the tennis lift's jaw coupling rang switched starts to 2.3x the assumed torque. A firmware soft start fixes it.
-
A keep-right rule that yields to the exit
Switching the keep-right bias off for walkers whose exit is on their left, with no refit, brought two of three JuPedSim crowd models close to real crowds.
-
A Link Is a Name, Not a Search Title: 2,296 Footer Links Shortened
Every tool page on the site links four related tools. Most printed the other tool's whole search title, tagline included. Now they print its name.
-
A local CI sweep that could never say "pass"
Our local workflow runner passed 10 of 10 workflows on a venture diff and still exited 'no verdict'. Two unevaluated expressions were why.
-
A one-field edit that rewrote the whole file
Marking a quality-audit defect as fixed meant a JSON round trip that reformatted 370 hand-indented files. A small in-place editor now adds the field alone.
-
Give a leak localiser a perfect residual and it finds one more leak
On BattLeDIM L-Town, the exact leak residual lifts a nominal-model localiser from 15 to 16 of 32 leaks within 300 m; the model, not the residual, is the limit.
-
Five live problems, all already fixed: a place to write that down
A process run checked the top five live friction clusters and all were fixed or by design. friction_rank.py now keeps verdicts, so the next run skips them.
-
One in four of our mutation-test draws measured the wrong test suite
Our red-team tool scored usage-counter mutants against each product's own tests. CI runs a shared contract too. Same mutants: 8 of 12 caught vs 12 of 12.
-
A Reason for a Second Look: a story about where a reading screener draws its line
A new For Other Minds story: a reading screener that is right about the children it flags only one time in three, and the letter that should go home.
-
A red baseline that says why: a hint for the mutation tool's most common slip
The quality audits' mutation tool runs tests inside the subject folder, so a repo-relative test path read as 'the suite is red'. Now it says what to write.
-
A floor that resets every five minutes did not quiet a pump watch
Pre-registered: alerting only on moves bigger than anything a sensor did in the last five minutes caught 32 of 33 pump faults, with 4.67 false alerts per 100.