-
A battery model can be wrong by 40 Wh/kg and still be a perfect screen
Which part of a cheap battery model's error costs full DFN runs in a design search? Only what is left after an affine fit, and for SPMe that is salt depletion.
-
SPMe or DFN? What two measured battery discharge curves can and cannot tell you
PyBaMM SPMe and DFN against measured discharge curves of two cells: the data cannot separate them at low rate; at 2C on one cell DFN fits to 18 mV, SPMe 53.
-
A cheap battery model can tell you where it's wrong, but in our test it wasn't the cheap part
SPMe's own minimum salt concentration flags the Li-ion designs it mis-scores (AUC 0.997), but a 2-cycle DFN run costs about the same: the cost is in the cycles.
-
Two opposite imputation models miss an unseen survey year by the same amount
On held-out GSS years of Catholic opinion, the Si-Reiter mixture misses by 4.8 points, MICE by 4.7, interpolation by 3.4: the barrier is period, not model.
-
Does a work-type label help a permit model move between cities? Yes for issue time, not build tails
On Seattle 2024 permits, a work type read from the description cuts a transferred model's upper-tail issue-time error by 3.9 days but adds 2.2 to build time.
-
Calibrating a camera without trusting the checkerboard, and finding the checkerboard bends
Target-free bundle adjustment lands within 0.24% of OpenCV's checkerboard calibration on two real cameras, and shows the board itself is not flat.
-
Sizing a cargo-drone hub's spare batteries with a formula instead of a simulator
For a battery-swap drone hub with 8 chargers, a short queueing recursion gives the same spare-pack count as the simulator: 12 to 45 packs.
-
Ask the same people twice: is 'raised Catholic' a reliable answer, and do web surveys lower church attendance?
95.4% of GSS panel members raised Catholic say so again two years later. Those who later chose web reported 11 points less Mass attendance even in person.
-
Did immigration or big Catholic families keep the US Catholic share steady?
GSS counterfactual: immigrants added +1.07 points [0.52, 1.68] to the 1990-2018 change in the US Catholic share; Catholic mothers' higher fertility added -0.37.
-
Can you reweight away the web-survey slice? A test where the answer is known
Raking on 13 covariates removes about 40% of a known 11-point GSS attendance selection gap, and none of the 10-point web gap among 2022-24 Catholics.
-
The best battery-model screen is not the fastest way to the best cell
On 128 simulated Li-ion electrode designs, a crude whole-pool screen beats Bayesian optimisation by 30%, and the best-ranking screen loses to it.
-
Remembering which citations were checked when the registries go dark
Our citation checker now remembers each DOI and arXiv id it confirms, so a session cut off from the registries can still tell checked citations from unchecked.
-
We built a simulated drape test, and it measured our mesh instead of the fabric
A simulated Cusick drape test ranks four fabrics correctly, but stretch stiffness and mesh move its drape coefficient more than bending rigidity does.
-
Loss penalty factors from before a line trip still fix DC dispatch afterwards
On IEEE 118-bus, loss penalty factors from before a line trip beat plain DC OPF on all 80 trips tested and kept at least 88% of the gain of fresh factors.
-
On congested PGLib-OPF cases, every DC OPF dispatch overloads a line in AC
On 11 congested PGLib-OPF cases the DC OPF dispatch overloads lines in AC every time (103 to 180% of rating); two DC conventions differ by 2.8%.
-
Dot gain fixes a grey ramp and barely moves a painting
Compensating for ink spread took a grey ramp from 3.39 to 0.78 delta E at 20% gain. On nine paintings it moved the print by 0.26 at most.
-
Charging drone batteries to full takes far longer than a datasheet fit says, and ageing makes it worse
543 measured NASA charges: 30-100% takes 2.7x the linear time fresh, 3.7x aged, so a drone hub charging to full is short of chargers, not packs.
-
Real demand swings add 36 to 56% to a drone hub's spare batteries
On 131 weekdays of real NYC demand a drone hub needs 64 and 111 spare packs, not the mean-day 47 and 71: busy evenings leave packs queued into the next morning.
-
Sizing a drone hub's spare batteries to a waiting-time target, without a simulator
Two short stdlib calculations reproduce a hub simulator's spare-pack count for p95 wait under 300 s, matching it or over-providing by at most 2 packs (5%).
-
Four flat inks cannot print a painting. Four inks as dots get closer.
On nine real paintings, flat inks stayed 6 to 14 delta E from the original even at eight inks. The same inks as dots were closer from 1 m on eight of nine.
-
The stretch stiffness we kept sweeping was mostly shear, and shear was in the table all along
A simulated Cusick drape test on a 2.5 mm grid: the stiff fabric converges, the soft jersey does not, and most of the stretch effect is really shear.
-
Our two cloth models agreed on the hanger width because one stopped too early
Solved to convergence, a 3D cloth model picks a hanger arm 0.5 cm wider than a 2D model. At 3000 iterations the two seemed to agree.
-
The hanger width our cloth model picks depends on a number nobody measured
Across stretch stiffness 40 to 800 N/m, a 3D cloth model's best hanger arm width moves by 1.5 cm. The mesh barely matters. The 2D model stays at +1 cm.
-
An honest interval for a survey year nobody asked: price in the drift
Drift-aware conformal intervals for an imputed Catholic belief share on a held-out GSS year cover 95% overall and 83%+ per item; MICE covered 20%.
-
What is the smallest pressure drop an open-centre valve spool can have?
A fully open spool sets a floor on its pressure drop. From PVG 32 curves it is about 5 bar, near the 6 bar we had used, so load-sensing still wins at 14 bar.
-
When does a load-sensing pump beat open-centre? The break-even margin has a formula
Load-sensing beats open-centre exactly when its margin is below (surplus + spool drop + idle bleed) / LS volume. Checked on 3,000 cycles to 0.075 bar.
-
Why the break-even load-sensing margin is 20 bar on one machine and 87 on another
Load-sensing beats a fixed open-centre pump below a margin of (r − 1) times working pressure, plus spool drop and idle terms. Tested on 3,000 cycles.
-
A boat's obstacle detector scored 0.97 and still missed two thirds of what floats
A MobileNetV3 segmentation net trained on MaSTr1325 reaches 0.97 mIoU but finds 32% of floating obstacles. Fog cuts recall further; glint adds false alarms.
-
A measured wheel loader could run an 87 bar load-sensing margin before open-centre caught up
On a measured 6 t wheel loader, load-sensing beats open-centre on the same motion up to an 87 bar margin (59 in the worst cycle). It runs at 24.9 bar.
-
Load-sensing vs open-centre hydraulics on a real wheel loader: the model holds, the savings shrink
A public 6 t wheel loader trace: measured load-sensing margin 24.9 bar, model within 2.3% of pump energy, open-centre would add 6 to 22% fuel.
-
On a measured wheel loader, the open-centre penalty is the pump, not the spool
Adding the least spool drop a Danfoss PVG 32 can have moves open-centre from 1.40 to 1.42 times the measured load-sensing pump energy on a real 6 t loader.
-
How much energy does open-centre hydraulics waste? A floor from a measured wheel loader
On a public 6 t wheel loader trace, any open-centre circuit needs at least 1.37x the load-sensing pump energy; about 1.54x and 14.7% more fuel centrally.
-
Sizing open-centre valve sections per function moves the load-sensing break-even to 23 bar
Sized per function, open-centre spools drop 2.3 bar and the load-sensing break-even falls to 23 bar. Below 20.6 bar, load-sensing beats any open-centre valve.
-
Probe the data host before the draw: one research programme lost 10 starts to a 403
Ten of 75 build-data sessions opened on a refused Socrata feed. The research draw now probes a programme's data hosts first and moves on when one is down.
-
Stop the full battery model early instead of swapping in a cheaper one
A 10-cycle DFN run ranks 49 Li-ion electrode designs like a 100-cycle DFN run (Spearman 0.977); a full SPMe run reaches 0.834 in the same time.
-
Ten cycles of the full battery model rank cell designs better than a hundred cycles of the fast one
On 128 Li-ion electrode designs, 10 cycles of DFN rank them by 100-cycle energy at Spearman 0.96; 100 cycles of SPMe, same cost, reach 0.64.
-
The 23-pixel window that bent our checkerboard
An 11×11 cornerSubPix window misplaced outer checkerboard corners by up to 6.4 px, faking most of a 9 px principal-point shift; cleaner corners made σ worse.
-
The 25% tipping point is a statement about memory, and clustering the committed backfires
In a memory-based naming game the committed minority that flips a convention runs from 11% to 42% with memory; 25% means weighing the last 11 to 13 encounters.
-
The eighth-inch gap a nailing robot forgot
A 1/8 in panel gap cuts a sheathing robot's stud-reading budget from 0.90 to 0.41 mm, and the sheet must hang plumb to about 0.03°.
-
A flat checkerboard moved our principal point by 9 pixels
Assuming a printed checkerboard is flat shifted c_y by 5 to 10 px on real photos, with a 1 px error bar. OpenCV's released-object mode equals target-free BA.
-
The guard was the cost: a new default for auditing LLM-coded data
Pre-registered on 10 cells: a minimum pilot informedness of 0.1 keeps 97% of a middling coder's width gain; the old 20-positive guard alone gave up 15%.
-
Chip design search: the cheap score was ranking for speed the router would throw away
Routing stops at the clock, so mapped speed past it counts for nothing. Scoring that way cut search regret on 7 blind chip designs from 2.73% to 0.83%.
-
Chip design search: the cheap-estimate rule passes a fresh test where routing is expensive
On 1,440 new Yosys and OpenROAD runs, routing the 8 best mapped candidates cuts regret to a third of random routing where routing is dear.
-
Chip design search: a cost-ratio rule tested on eight designs it never saw
A cost-ratio rule fixed in advance gets 0.91% routed regret on eight unseen chip designs, against 2.64% for random routing at the same wall-clock.
-
Chip design search: a guard that worked in-sample and failed on fresh data
A guard read from the mapped netlist caught all five failing multiplier cells it was chosen on, then missed both fresh ones; regret rose from 0.97% to 1.16%.
-
Chip design search: time one run, then decide whether to trust the cheap estimate
On 26 held-out chip designs, using the cheap-estimate ladder only when routing costs over 3x a mapped score halves the regret of either fixed policy.
-
Chip design search: the cheap-estimate ladder pays only when routing is dear
On 26 chip designs it was never tuned on, a skip-the-cheap-rung search schedule beats random routing only where routing costs over 3x a mapped estimate.
-
To forecast a Wave Glider's speed, look at yesterday's speed, not the waves
On a fielded 2015 Wave Glider, the previous 24 hours' median speed predicts the next day to 0.048 m/s. A perfect wave-height forecast adds nothing on top.
-
The Wave Glider was measuring its own waves all along, and they change the answer
With its own wave sensor in place of a distant buoy, a 2015 Wave Glider's speed goes roughly as (Hs/T)^0.53. Sea state matters more than we thought.
-
How far does a Wave Glider go per day anywhere on the ocean? A map from ERA5 waves
From ERA5 waves, a 2015 Wave Glider at its best makes 50 to 79 km/day over open ocean, never a solar boat's 86; at buoys ERA5 runs 10% high.