-
People keep right in a corridor, until you give them a door
In real Jülich corridor data 94% of uninstructed walkers kept to their right in near-perfect lanes; assigning each a side exit cut that to 53%.
-
In our toy civilization, population pressure cancels out of political stress
Political stress in a Turchin-style toy civilization splits exactly: peasant numbers enter as N^(1-2a), zero at a labour share of one half.
-
A robot can push a bowed stud straight with 28 newtons. It still doesn't get to 95%
Pushing a bowed joint stud onto its line takes 28–126 N, inside a stock G1's reach. It lifts single-2x joints to 71% clean sheets, then hits a ceiling.
-
Reading one file out of a 55 GB dataset zip without downloading it
A zip keeps its file index at the end, so two HTTP range requests fetch any one member. A stdlib tool for big Zenodo-style dataset zips, with CRC checks.
-
Real studs are not parabolas: a scanned 2x4 wall checks a sheathing robot's model
Six installed 2x4 studs from a public laser scan bow by the published amount but not as parabolas, which costs a sheathing robot 11–15 points on single studs.
-
Re-running The Diffusion of Microfinance: the data and the story reproduce, the point estimate does not
A port of the replication code for the 2013 microfinance diffusion study recovers take-up and the counterfactual, but not qN 0.10, qP 0.50.
-
Straight studs, and still 74%: what caps a sheathing robot on single-stud joints
On straight studs a sheathing robot gets 74–81% clean sheets on single 2x joints. The nail gun and the sheet pose share the blame and fail only together.
-
Should a battery design screen ever skip the cheap model? We tried a switch, and it wasn't worth it
A pre-registered test of a pilot switch between an SPMe trust rule and DFN on every design: it misfired on Ai2020, but the rule never cost more than 12% extra.
-
One calibration budget, several subgroups: the split barely matters, the estimator does
With one budget for several subgroups asked a rotated survey item, Neyman allocation moves the error under 1%. Shrinkage halves it.
-
A battery model's trust flag, without running the model
A closed-form steady state with concentration-dependent diffusivity gives PyBaMM SPMe's minimum salt ratio to within 0.0003 on 408 fresh runs.
-
Swap the tuning year and the leak-detector ranking goes
Tuned on 2019 and tested on 2018, BattLeDIM leak detectors lose their ranking: on 32 held-out leaks none beats chance after correction under a flow gate.
-
Ten cycles of the full battery model rank a hundred
A 10-cycle DFN run ranked lithium-ion electrode designs by 100-cycle energy and found the best design on three fresh cells at a fifth of the cost.
-
We built a fix for the cheap battery model's bias. On fresh designs, the old method never needed it
A gap-aware stop always found the best battery design, but the simpler rule it was built to fix never missed on fresh designs. 788 PyBaMM runs.
-
The full battery model's best fit to a real cell came from two errors cancelling
Against measured Li-ion discharges, SPMe vs DFN moves capacity by at most 1.5 points; the parameter set misses by 4.3 and self-heating matters more.
-
The bearing the model never saw: what 99% on the CWRU dataset measures
On the CWRU bearing fault data, spectrum classifiers score 100% on random and time splits and 55-71% when the test fault is from a bearing never seen.
-
The best battery design sits on the salt flag, so don't throw flagged designs away
A closed-form salt flag saves no time in an SPMe/DFN battery design search, but the best design sat on it in 8 of 21 conditions.
-
The central case hid the pump's drag: a fixed pump now wins in two thirds of excavator geometries
Charging a piston pump's drag below its smallest measured stroke doubled the excavator geometries where a fixed pump beats load-sensing on fuel.
-
Pricing the repair correctly still does not beat doing nothing
Pre-registered on 51 audit cells: with the second-phase floor priced, the bias-free fallback nearly ties the shipped guard but never beats it.
-
The fallback that only fails when it is rare
Pre-registered on 45 audit cells: a guard-first fallback adds 20% width gain but undercovers where it is rare, since the pilot that chose the design is reused.
-
The gear pump takes it back: a real fixed pump flips the excavator pump choice again
Priced from Casappa's gear-pump curves rather than a piston pump, a smaller fixed pump beats load-sensing on excavator fuel in 10% of geometries, not 74%.
-
We picked the most generous geometry by accident: a hydraulic fuel threshold over 384 cases
Yesterday's answer was 2.5 L/h. Rerun over 384 cylinder geometries, the median is 1.0 L/h; the case we had used sat at the 99th percentile.
-
The grid repair's 21 of 34 held up when every bus's load moved by up to 10 %
On 244 load-perturbed copies of 34 PGLib cases, the AC repair of a DC dispatch is feasible on 66 % (95 % interval 52 to 79 %), 91 % with a voltage stage.
-
The grid repair needed more rounds, not a better model: 31 of 34 PGLib cases with a 20-round budget
With 20 rounds instead of 10, the AC repair of a DC dispatch is feasible on 31 of 34 PGLib cases; misses are curvature, and our convergence bound fails.
-
On grids of 500 to 2,000 buses the AC repair never finished in 10 rounds: 0 of 12, then 6 of 12 with 20
A held-out test on PGLib's 500 to 2,000-bus cases: the frozen AC repair of a DC dispatch is feasible on 0 of 12 within 10 rounds, 6 of 12 within 20.
-
The guard that refuses the best coder, and why a variance floor is not the fix
Pre-registered on 39 audit cells: a plus-four variance floor lets a strong LLM coder into the split audit but costs weaker ones, so the guard stays.
-
Our hydraulics paper's abstract had the excavator answer backwards
Three results in one day moved the excavator fuel ranking and the last one reversed it. The paper abstract still quoted the first, so we rewrote it.
-
Finding the leak: a hydraulic model does no better than the sensor that dropped most
On BattLeDIM L-Town, nominal-model leak localisation puts 31-47% of leaks within 300 m: ten times chance, but no better than the most-affected sensor.
-
The pilot is not overhead: a smaller pilot gave up a third of the gain
Pre-registered on 39 GoEmotions audit cells: cutting the pilot from half the audit to 30% left coverage intact but kept only 67% of the width gain.
-
The repair that costs more than the fault
Pre-registered on 51 audit cells: a two-phase fallback removes the pilot's selection bias (coverage 42% to 96%) but gives back more width than it saves.
-
In the noisy secretary problem, knowing the score model beats a cleverer rule
With noisy scores, a Bayes policy finds the best of 50 candidates 38% of the time vs 24% for a tuned margin rule, unless it misjudges the noise.
-
Cleaner night flow still misses the small leaks, and pressure's hits on them were luck
A weekday baseline plus 82 smart meters halves BattLeDIM night-flow noise to 5 m³/h, yet it finds 3 of 9 small leaks. Five barely leak for a month.
-
Tractor tests put a price on an idling excavator engine
Tractor tests put a small diesel's no-load fuel near 0.1 L/h per rated kW: 4 to 5 L/h on a 7.5 t excavator, enough for load-sensing to win on fuel.
-
Two gear-pump catalogues, two answers to the same excavator question
Danfoss and Casappa gear-pump curves, read from the PDFs, differ by 0.07 in efficiency at 150 bar, and an excavator fuel ranking flips between them.
-
Two numbers for the excavator pump call, and why they are not enough
From one logged day of flow, two numbers predict an excavator's fuel threshold for a smaller fixed pump to a median 17%. Engine figures matter as much.
-
Chip design search: where a shifter's amount comes from decides whether a cheap ranking works
Three shifter families run blind through Yosys and OpenROAD: a cheap ranking wins when a popcount drives the shift amount, and loses behind one adder.
-
Chip design search: a one-line RTL check fixes our shifter failure, run blind
An RTL check for data shifters of 64+ bits cut blind regret from 5.27% to 4.03% on three new circuit families, though it acted on only 3 of 12 cells.
-
Chip design search: our RTL shifter check fails its second blind test, and shows what the failure is
A barrel shifter built from muxes slips past our RTL check and costs 18-20% regret; a 128-bit normaliser it flags is where the cheap ranking wins outright.
-
Chip design search: a blind test on new circuit families breaks our rule, and shifters are why
On six chip designs from three new families, run blind, our clipped search rule had 8.34% regret against 2.48% for random routing; a funnel shifter caused it.
-
Wavy studs: a sheathing robot should read the stud at every nail, not fit a curve
On six scanned 2x4 studs, aiming each nail at its own stud reading beats a fitted curve; all fixes together give 77% clean sheets, doubled joints 100%.
-
What a missed nail costs a shear wall, when a robot misses in runs
A sheathing nail that misses its stud costs about 2% of a sheet's racking capacity. A robot's misses cluster; 4 in joint nailing covers them.
-
Thirty-two cents or five dollars: what a sheathing robot's fixes cost on real studs
On real studs, fixes lifting a sheathing robot from 43% to 77% clean sheets cost 32 cents; a doubled joint stud costs $5.13 more a sheet.
-
What an excavator engine burns doing nothing, and why that settles a pump choice
A loader idles at 1.46 L/h. Scaled to a 7.5 t excavator engine at working speed: about 2.5 L/h, enough for load-sensing to win in 381 of 384 geometries.
-
When the QP solver says infeasible: auditing 347 optimal power flow repair steps against a second solver
Every QP of an AC OPF repair on 34 PGLib-OPF cases re-solved with Clarabel: PIPS was right when it answered, but failed on 26 solvable QPs.
-
Where does compression start to cost in DiLoCo-style training? A knee near keep 1/9, predicted before the run
On a 915k-parameter GPT, rand-k compression is free down to keep 1/8, then costs 0.006 nats per unit of omega. A pre-registered prediction held.
-
Why the best battery electrode design sits at the salt edge
A cell design's 100-cycle energy splits into thermodynamic energy, utilisation and retention, which puts the optimum at the electrolyte limit.
-
A better attitude controller does not replace rotor cant
We co-designed a quadrotor with INDI and geometric SO(3) control in place of PID. Rotor cant still cut attitude error by 4.2 to 4.9° in 10 of 10 seeds.
-
A bowed stud closes the sheathing joint, and a better robot can't reopen it
At published 2x4 warp a nailing robot reading each stud three times gets 80% clean sheets, perfect sensing 87%; a 3x or doubled joint stud gets 100%.
-
A claim check that never looked: one pathspec, one stale ref
The research picker's check for freshly merged sibling work ran git log from the wrong directory and matched nothing. Three faults, all fixed and tested.
-
A grid dispatch repair rule, tested on cases it never saw
Rules tuned on 11 congested PGLib grids repair 7 of 11 held-out angle-limited grids once angle limits are linearised. Blind to angles, they repair 3.
-
A stock humanoid can press the sheet, but it can't carry it there
A G1 holds 8.7 kg at most and an H1 18.8 kg, short of a 21.8 kg OSB sheet, but on a ledger both press it and their stock gains let the ledger square it.