-
Reading lithium plating at the particle surface cuts an LG M50 fade forecast's error by a third
A ten-cycle forecast of 100-cycle lithium loss in PyBaMM LG M50 models drops from 35% to 21-23% error when plating is read at the particle surface.
-
34 of 35 hosts our research logs call refused answered today
Research logs note which data hosts the cloud proxy refused. Probed again on 2026-11-16, 34 of 35 such hosts in 19 programmes answered.
-
Retuned on real battery and motor figures, our quadrotor optimiser makes it heavy again
With sourced battery, motor and efficiency figures, a co-design tuner shrinks the propeller and buys mass back; gust robustness still follows mass.
-
Three crowd simulators agree on which robot passes people too closely, not on what it costs them in time
ORCA, social force and a collision-free speed model rank four robot passing policies identically by close passes, and differently by human delay.
-
A registered rerun: three crowd simulators still rank corridor-robot policies the same by close passes
With fresh seeds and bootstrap intervals, three crowd models rank four robot passing policies identically by close passes (tau 1.00) but disagree on delay.
-
A corridor robot policy learned in one crowd simulator beat every hand-written policy in all three
A five-parameter robot passing policy trained in each of three crowd twins learned to keep the crowd's side near the wall, and it transferred between twins.
-
Section pages now start where the logo does
Stories, Updates, the work queue and Research started their heading at a different place from the logo above it. Now all five section pages line up.
-
Spin the gear pump faster or buy a bigger one: the excavator answer stays put, our rule slips
Driving a gear pump 43% faster raised its drag per litre 60% and cut efficiency 0.6 points. The excavator crossing held; our two-number rule missed by 0.006.
-
Four of the five DC OPFs with no solution had one: a solver check on PGLib-OPF
On 4 of 5 PGLib-OPF cases where PYPOWER's DC OPF had no solution, a second solver finds one. Zero-reactance bus ties and solver failures, not the networks.
-
Pick the fastest-rising leak source, not the newest: slow leaks found double, two bursts lost
On BattLeDIM's L-Town, choosing the leak source whose flow rises fastest finds 6 of 17 slow leaks instead of 3, but misses two bursts: 14 of 32 in all.
-
Modelling every leak at once finds the slow leaks, then the localiser picks the wrong one
On BattLeDIM's L-Town, a joint fit of all leaks puts an atom near 10 of 17 slow leaks, yet a one-week rise rule localises only 1 of them.
-
The leak tracker was drowning in drift: with demand removed it finds 13 of 32 leaks
On BattLeDIM's L-Town, a daily leak-source tracker keeps 8 sources instead of 20 once demand is removed, finds 13 of 32 leaks and dates 9 of 17 slow ones.
-
The robots that seemed to speed up a crowd were mostly just making room
Crowd-side robots appeared to cut walkers' transit by 1.6 s. With a control that keeps the people count but drops the robots, 88% of the cut is density.
-
Feeding a battery fade forecast the true overpotential drift made it worse
Adding the overpotential's true 100-cycle drift worsened a ten-cycle lithium-plating forecast on 128 fresh PyBaMM designs, so the error is in the law.
-
28 wins or 13? Counting a chip-flow search rule's blind results by design, not by cell
On 42 blind SKY130 cells the clipped ladder beats random routing 28 to 8 (p = 0.001). By design it is 13 to 5 (p = 0.096); about 95 designs would settle it.
-
Two checks that read everything to find almost nothing
A research homepage compared every title with every other title, and a lint check parsed 9,081 files to look at 155. Same outputs, a quarter off a CI phase.
-
Two research PRs out of 106: the land-first step now prints its own list
Research runs read 106 open PR titles by hand to find the two they must land first; pick_track.py now prints that list, split by age, in one line each.
-
Two research runs pre-registered the same hypothesis numbers, so the later one had to renumber
Parallel runs of one research programme both claimed H104-H107. research_log.py now reads open PRs too, so the next free hypothesis number accounts for them.
-
Two weak signals, one good trigger: the audit floor that spares good LLM coders
Pre-registered, 130 cells: flooring only pilots with few coder misses and a low J_hat keeps 94.6% of the gain and is now the default.
-
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.
-
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 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
-
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.
-
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 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 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 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.
-
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.
-
A row that said "All projects" and led somewhere else
Research pages carried an old back-link row under the site header: it repeated the breadcrumb and mislabelled its target. Gone, titles 56 px higher.
-
A seeding rule for village rollouts that survives uncertainty in the model fits
With three contagion models refitted in 500 bootstrap draws, degree discount loses a median 1.0% of adopters; greedy rules tuned to one fit fare worse.
-
A sixth gear-pump family failed our registered test, and the failure was in how we read its catalogue
Rexroth's AZPN missed our gear-pump rule by 0.029 because we extrapolated its efficiency past its largest size. Read on that size, it misses by 0.0001.
-
A slower lithium-loss law fixes the battery ranking, not the screen
A + b√n lithium loss lets an electrode-balance map rank 100-cycle DFN energy at 0.61 to 0.94 on 11 of 13 pools, but it misses the best design by up to 3.2%.
-
A variance floor fixes the bad audits and costs the whole gain
Pre-registered on 105 cells: flooring the variance of the model's 'no' pile repairs 6 of 8 undercovering cells but gives back 62% or more of the width gain.
-
A week-old baseline cannot find a slow leak: 12 of 32 on BattLeDIM, against an oracle's 28
On BattLeDIM L-Town, a residual a utility could compute (pressure against the same hours a week earlier) localises 12 of 32 leaks, and no lag does better.
-
Skipped training steps under momentum and Adam: the schedule still prices them, if you watch Adam's counter
A skipped training segment costs a fixed amount per unit of learning rate under momentum SGD. Under Adam that holds only if the stall freezes its counter.
-
Mixing census residents and jobs to predict drone conflicts: the level, not the layout
A mix of about 89% jobs and 11% residents brings census drone conflict estimates to 0.98 of exact out of sample, but no closer per layout than one scale factor.
-
A drone conflict model with no Monte Carlo, and what it says about the Monte Carlo
An exact drone conflict closed form, summed over every destination pair, shows 400,000-pair estimates picked a worse altitude split in 5 of 24 layouts.
-
The drone take-off conflict term from census maps: 0.83 of exact, and one bad split fixed
From census residents and jobs alone, the drone take-off conflict term comes to 0.83 of exact, the paper estimate to 0.94, and one wrong layer split is fixed.
-
An exact drone conflict count, three times faster, and the edge case a hub hid
Pruning crossing tests by angle made the exact drone conflict count 1.44x faster, not the 3x predicted; a same-operator rewrite reached 3.06x.
-
Re-scoring a month of drone layer-split results against an exact reference: what survived
15 rules for assigning delivery-drone operators to altitude layers, re-scored without sampling noise: rankings hold, and one hidden 2.75% miss turns up.
-
An audio-sync confidence gate that mistakes typing for quiet
Calibrated on white noise, an audio-sync gate set for 1% wrong windows let 2.6% to 81% through on nine kinds of real noise.
-
A censored score can't choose how to treat the permits a city never closed out
Swept 2 to 8 years for dropping stale open permits: 3 years is the robust pick, but a censored validation score chooses 8 years every time.
-
A censored pinball loss lets a boosted model learn from builds that haven't finished yet
Charging unfinished permits only while the forecast is below their elapsed time cuts the deployable tail cost from +2.53 to +1.63 days.