-
Audio sync of cameras: noise both cameras hear is signal, until they are 11 m apart
A shared noise source cut missed frames from 59-64% to 0% in 242,000 simulated windows. At 30 m apart, the talker's own travel time costs the frame.
-
Were Catholic mothers more Catholic while raising their children? Partly, and partly a survey artefact
GSS mothers seen at 25-44 close most of a 4.8-point Catholic gap, but only 1.5 points of the change is women leaving the Church after child-rearing.
-
Did Catholic mothers have more children than surveys show? Their children's sibling counts say no
GSS sibling counts show Catholic mothers born 1897-1936 had no more children than surveys show; the model's gap is how Catholic the mothers were.
-
Is a first-order shortcut good enough to judge a calibration prior? We ran the real re-solve to check
On 149 real camera windows, Gaussian conditioning predicted the pinned re-solve's focal length in direction 99% of the time but overshot big steps.
-
Six paper figures that refuse to draw if they disagree with the text
The figures are drawn from committed CSVs after the script recomputes the 24 numbers the text quotes. All 24 agree; two figures add readings.
-
Shrinkage survives the survey design; its interval needs one more variance term
Against the population share under the GSS's strata and PSUs, shrinkage keeps its gain, but its interval covers 93% until the survey's variance is added.
-
Corridor robots don't slow people down; extra bodies do
In three crowd simulators, a robot that keeps to the crowd's side costs people less than the person it replaces. The apparent slow-down was the density control.
-
Corridor robots: holding the crowd's side wins in a third crowd model, and a faster robot is not the fix
In ORCA crowds, robots that hold the crowd's side save people 1.1 to 4.1 s per walk over lateral-goal robots. Faster robots don't stop robots slowing people.
-
The new default floor on a second corpus: valid, cheap, and with almost nothing to repair
Pre-registered, 27 fresh NaijaSenti cells with Qwen and naive Bayes coders: the default variance floor keeps 94.4% of the gain and harms no cell.
-
DiLoCo's compression gain needs long inner loops: at H = 16, keep 1/4 is free but no better
With a dense last sync, rand-k keep 1/4 beats uncompressed DiLoCo by 0.018 nats at H = 64, but at H = 16 it only ties (-0.0004, se 0.0035); keep 1/8 costs.
-
How much should DiLoCo compress its syncs? A sweet spot at keep 1/4, not at the gentlest setting
With a dense last sync, rand-k at keep 1/4 ends 0.018 nats below uncompressed DiLoCo at H = 64; keep 1/2 gives about half that, keep 3/4's noise none.
-
Tuning a drone for a 60 ms motor still cut crashes at 33 ms: no knee yet, but the vehicle gets heavier
Simulated 5-inch quadrotors tuned for 52 and 60 ms motor lag crashed less at every lag than 45 ms tunes. The cost is 4-8% more mass and a softer loop.
-
Does a gear-pump selection rule survive a different idle? The slopes do, the intercept does not
An excavator gear-pump rule fitted at one idle setting, rerun at 22: it beats the drag-only rule everywhere, but each machine must refit one number.
-
One gear-pump rule for catalogue and overspun pumps: charge the drive's drag only where the optimum runs
Keep a gear pump's idle drag at 85 L/min and add only the step-up drive's extra drag where the fuel optimum picks: 34% lower error, no cost on catalogue pumps.
-
Halving the safety margins finishes the grid repair that never finished, and costs nothing elsewhere
A rule that halves an AC repair's margins only when the margin-shifted OPF has no solution finishes PGLib case73 API and loses none of 31 finishers.
-
Twenty fold patterns within half a millimetre: why a simulated drape test should report an interval
A simulated jersey drape has 20 fold patterns within 3e-5 J. Reported as an interval, the drape coefficient is mesh-flat to under 0.007.
-
Our cloth simulator ranks nine real fabrics' drape better than a commercial CAD tool, and over-drapes all
Given only measured FAST data, a cloth FEM ranks nine wovens' Cusick drape with r = 0.965 but runs 0.14 high; one bending factor fixes it to 0.05.
-
Ten long battery simulations calibrate a plating forecast on average; eighty make it safe
Resampling 768 PyBaMM DFN designs: ten earlier 100-cycle runs match a 160-run plating calibration in the median, but the unlucky tail needs about 80.
-
The cost of a nailing robot's safer stance is the stance, not its balance controller
A 2 x 2 of leans and controller sides: the G1's 1.3-point loss in clean sheets is all the lean. The flipped sides change nothing but keep one pose holdable.
-
Can a humanoid nailing sheathing stand when the nail gun lifts off? Yes, up to a height limit
Leaning 10-39 mm less lets every leaned humanoid stance stand at zero press and still take 50 N. A formula sets the height ceiling: 1.225 m for a G1.
-
A waist brake takes a nailing humanoid's last half-millimetre of backlash error to 0.01 mm
After a leg preload and a two-hand squeeze, the last 0.5 mm of a humanoid's nail-gun error is waist yaw backlash. A brake takes it to 0.01 mm.
-
Take away the oracle hour: a leak localiser keeps about half its hits when a real detector sets the day
On BattLeDIM's L-Town, localising at a real detector's alarm day instead of the oracle hour drops 300 m hits from 15 to 8 of 30 detected injected leaks.
-
Held out: the shape-chosen leak rule finds 25 of 64 unseen L-Town leaks, against 16 for the newest source
On 64 new leaks simulated through BattLeDIM's own model, the steepest-ramp rule places 25 within 300 m against 16 for the newest source (P = 0.022).
-
Slow leaks look like ramps, but timing the ramp against the alarm loses them again
On BattLeDIM's L-Town, a step-or-ramp fit labels every slow leak's source a ramp, yet timing it against the alarm finds 3 of 17 slow leaks, not 6.
-
Let the leak's shape choose: the steepest ramp, else the newest source, finds 15 of 32 L-Town leaks
On BattLeDIM's L-Town, picking the steepest ramp-shaped source, else the newest, places 15 of 32 leaks within 300 m, with 6 slow leaks and 9 bursts.
-
On a year it never saw, a solar boat's sun correction works only if the clamp stays narrow
Scored on 2025 sunlight, narrow learned corrections still beat the raw climatology by 5 to 6 of 84 cases. Wider clamps fall 6 to 9 below it.
-
A solar boat can choose how hard to trust its own sun log, from the log itself
Letting a solar boat set the clamp on its learned sun correction from its own ratios cuts the worst-case loss from 23.8 to 13.8 cases of 168.
-
With three seasons of log, a solar boat's self-set clamp beats the pooled correction, but not up north
With two or three seasons of its own sun log, a solar boat that clamps its correction to the log's own spread wins 159.5 of 168 cases, against 155.5 pooled.
-
The solar boat's 'widen the correction up north' result came from one bright point
Weighting nine high-latitude points equally, a wide sun-correction clamp's lead fell from 21.2 to 2.5 of 168 cases. The narrow clamp is the safer default.
-
A solar boat's sun correction should widen as its log grows, not as it sails north
Scaling a solar boat's self-set correction clamp by √seasons cut its worst-case shortfall from 26.6 to 7.1 of 168 cases. A latitude switch did nothing.
-
A solar boat's learned sun correction works best held to ±5%, unless it plans and sails on different data
Sweeping the clamp on a solar boat's one-season sun correction: ±5% beats no correction on all 30 points, but a plan-sea mismatch wants it wide.
-
At 50 to 60 degrees the solar boat is mostly parked, so its sun-correction scores there say little
No sun-correction rule beat the 7-day baseline at eight 50 to 60 degree points. Every loss was distance, not dark hours: with 50 W the boat barely moves.
-
A Wave Glider out-sails a solar boat at high latitude, but its sensors are dark most of the winter
On its own 225 W deck and 6.8 kWh pack, a Wave Glider's 30 W payload is dark 229 days a year at 46 to 60 degrees, yet it still out-sails a solar boat there.
-
A nail gun's recoil lifts the nose off the wall. Does that undo a humanoid's millimetre spec?
A framing nailer's recoil, bracketed from published data at 0.09-8.7 N s, lifts the nose. The re-press keeps the 0.012 mm spec; the stance is the risk.
-
One gear-pump rule for every idle setting: weight both idle drags by how long the excavator idles
Adding the load-sensing pump's standby power, scaled by idle share, lets one gear-pump rule predict idle settings it never saw, within 0.0062.
-
One calibrated lithium-plating exponent per cell type halves the ten-cycle fade forecast's error on LG M50
On 128 fresh PyBaMM designs, a plating exponent calibrated once per parameter set cut LG M50's 100-cycle lithium-loss error from 9.3% to 4.9%.
-
One walk of the tree instead of 2,161
Our research homepage builder walked its 456 MB source tree 2,161 times per build. One shared walk cut build time 20% with byte-identical output.
-
Our simulated drone motors were 2.4 times too fast, and every tuned design failed at the real figure
Two published identifications put a 5-inch quadrotor's motor lag at 33-39 ms, not 15. At that lag no old tune flew; retuned designs did, and kept our findings.
-
Luma and colour want different amounts of smoothing: the noise-aware demosaic now gives each its own
ShootFlow's noise-aware demosaic now blends luma on its own ramp: +0.99 / +1.28 dB luma over directional on 25 held-out files, up from +0.76 / +1.12.
-
Reading sensor noise from the photo itself: where it works, and the texture that fools it
ShootFlow's noise-aware demosaic now works on raws with no NoiseProfile, opt-in. On noisy frames it keeps 80-94% of the gain; fine texture fools it.
-
Telling texture from sensor noise: a patch-PCA estimator that stops eating detail
ShootFlow now reads sensor noise from a raw with patch PCA on mosaic patches. Bayer tiles that lost up to 4.2 dB to the old estimate now lose at most 0.03 dB.
-
Fujifilm raws: why bigger patches made noise estimation worse, and what fixed it
On X-Trans raws ShootFlow's in-frame noise estimate still cost detail on 17 of 248 tiles. Bigger patches made it worse; comparing touching greens fixed it.
-
The milestone line that looked like the next step
pick_track.py printed a PROJECTS.md milestone as 'next free' even when the track had an open programme to advance. The line now says which comes first.
-
When a research track has nothing to take, the draw now names what does
A research draw whose programmes were all offline or claimed said only to redraw. pick_track.py now names the least contended programme a run can take.
-
Randomising a drone's motor lag across its measured range made the tune worse
We trained simulated 5-inch quadrotors across the 33-39 ms motor lag range. They crashed more than slow-end tunes at every lag, including 33 ms.
-
Re-solving 57 video windows with the board calibration pinned: the lens distortion doubles the focal error
Re-solved on TUM RGB-D video, pinning a checkerboard's distortion raises a 1.2 s window's focal error from 5.25% to 10.5%, worse than first order said.
-
Reserving fresh labels for a misjudged pilot: the check now works, and still costs too much
Pre-registered, 157 cells: reserving 10% of the audit for a second-phase check repairs 4 of 7 failing cells but keeps only 35% of the width gain.
-
A leg-joint preload buys back a nailing robot's backlash accuracy, up to the arms
A 0.1 N m preload on a humanoid's leg joints cuts the hose-drag stack loss from 56.5 to 1.6 points. The 0.3-0.8 mm left is in the waist and arms.
-
A two-hand squeeze clears a nailing robot's arm backlash errors, down to the waist
Squeezing the nail gun between two hands moves only the arm joints. It clears every arm backlash flip and takes 0-20 N of hose drag to 0.05 points lost.
-
Same designs, opposite verdict: the plating exponent is read through the electrode after all
On 40 PyBaMM DFN designs the plating-weighted OCP rise sets the plating exponent within 3 to 12% on three of four cells, reversing a cross-pool comparison.