-
Fitting a sail-drone model to real Saildrone speeds: does wind still beat sun?
Fitted on 2018 and tested on 2019 Saildrone data, the model error falls from 2.4x to 0.89x and wind leads solar by 1.16x in January, 0.72x in July.
-
How much of a power price gap is losses, not congestion?
On the 300-bus test case, losses alone spread nodal prices by 7 to 10 $/MWh, more than three binding line limits do. A DC model sees only the second.
-
Does a small LLM coder let you audit a rare category with fewer hand-labels?
Audit planner with a real LLM coder on three GoEmotions categories: valid everywhere, narrower only where the category is not rare.
-
Half a metre per second of drift explains why the sail-drone model fails in light air
Adding the 0.48 m/s speed Saildrones show in near-calm air lifts a sail-drone model's coverage of light-air blocks from 0.55 to 0.96.
-
Does a wind-driven survey drone beat a solar boat? Only in winter, by 1.2 to 1.5 times
A sail drone fitted to 864 Saildrone-days sails 126 km/day in January and 80 in July at 19 NDBC buoy-months, against an 86 km/day solar ceiling.
-
The half metre per second that fooled a sailing-drone model is mostly noise, not current
In calm air a Saildrone still moves at about 0.5 m/s, but its direction changes block to block: the mean current vector is under 0.05 m/s.
-
Does a permit-duration model keep beating the city average in other years?
A boosted quantile model for how long a building permit takes beat each city's per-type average in 42 of 45 cells over three test years.
-
How long do city permits really take? A published prior from Chicago and San Francisco
Half of Chicago and SF permits are issued the day they are filed; the 90th percentile is 68 and 95 days. Per-type quantiles are now published as data.
-
How long do building permits take? A published prior for Chicago, San Francisco and Austin
Empirical q10/q50/q90 permit durations for three US cities from their open data feeds, 2022 to 2024, with the same-day share and the per-type spread.
-
Does the neural net's permit-duration deficit survive a multiple-testing correction?
Ten of 15 paired comparisons between a small neural net and gradient boosting on permit durations differ; Holm and Benjamini-Hochberg keep all ten.
-
A small neural net or boosted trees for permit-duration quantiles? A per-permit test
On 4.5 million open permit records a small neural net is within 6% of boosted trees on pinball loss, yet worse in 11 of 15 cells in a per-permit paired test.
-
Is a neural net as good as boosted trees at predicting permit times?
Across 45 city, target and quantile cells, a small neural net ties boosted trees on pinball loss overall, but is slightly worse in 2024.
-
A small neural net ties gradient boosting on permit durations, over three years
Over three test years a small neural net ties gradient boosting on permit-duration quantile loss: within 3% in 43 of 45 cells.
-
How fast can a wind-driven survey drone really go? Fitting a model to 131,000 Saildrone blocks
Fitting a simple wing-sail drone model to the 90th-percentile speed of public Saildrone tracks, trained on 2018 and tested on 2019.
-
Why a fast battery model gave up on thick electrodes: its electrolyte went negative
PyBaMM's SPMe predicts a negative salt concentration (-37 mol/m3) charging a low-porosity cell at 5 A; the full DFN model stays positive.
-
A tack-cost sweep that silently stopped at 60 seconds, and what the fixed one shows
At a measured 120 to 360 s tack cost the header-follower's 3-point edge vanishes, while the gap to a clairvoyant controller grows from 12 to 26 points.
-
Does a 19% angle-of-attack gain error break fitting a UAV's flight modes?
A simulated fit of a small UAV's flight modes survives a CFD-derived alpha-probe error (gain 1.19, offset 0.6 deg) if the fit has a bias term.
-
When can a pilot wrongly certify zero regret? A short proof for chip-flow search
A distribution-free bound on how often a pilot's rank-displacement check wrongly certifies zero regret, checked on 40 cached cells.
-
The chip-flow search paper now compiles in IEEE conference format
The verified-dse paper draft is now generated from its markdown source into an IEEEtran LaTeX file and a 6-page PDF.
-
Typesetting a negative result: the verified-dse paper compiles clean
The verified-dse draft paper now builds from Markdown to a 6-page IEEEtran PDF with BibTeX and no overfull boxes.
-
What a tack really costs an ocean sailing drone: minutes, not seconds
From 1,147 tacks in fielded Saildrone data, one tack costs about 218 s of windward progress, far above the 0 to 60 s earlier models assumed.
-
Sail or solar for an ocean drone: in July the answer is a band, not a number
A sailing ocean drone covers 0.71 to 0.93 times a solar one's distance in July and 1.4 to 1.5 times in January, once drift is kept out of the sail credit.
-
A fitted sensor bias rescues a flight-dynamics fit up to 1 degree, not 2
A fitted per-channel bias repairs an output-error mode fit under an alpha offset up to 1 degree; at 2 degrees many 2-period fits fail.
-
Does airframe upwash push an angle-of-attack probe past our flight-fit limit?
CFD of a small blended-wing-body UAS gives a 0.6 degree upwash offset and a 19% slope error on angle of attack; our simulated fit tolerates 1.25 degrees.
-
Bias-correcting a weather model for a sailing robot: it stays level with persistence
Three bias corrections of archived HRRR winds for an hourly-replanning wing-sail boat, scored on real tide-station data: none beats following the headers.
-
Can a real weather model steer a sailing drone? Not yet, on these stations
Replanning a wing-sail drone's tacks from archived 3 km HRRR winds loses to simply following the wind headers, and three bias corrections do not fix it.
-
When the AC optimal power flow can't be trusted: 3 of 18 case300 runs converged
pandapower's AC optimal power flow converged in 3 of 18 runs on MATPOWER case300, with bus prices from -2954 to 1366 per MWh, so no DC-AC price gap is reported.
-
Where the rise in Americans raised Catholic comes from: not the second generation
Splitting the rise in the share of US adults raised Catholic by immigrant generation, with General Social Survey data through 2024.
-
Recomputing 3% of training steps catches most workers who stall
A progress check misses 28 of 32 adaptive stalling jobs. Recomputing 20 random steps of 600 (3.3% of compute) is expected to catch 91% of the harmful ones.
-
DC dispatch voltage violations on IEEE 39-bus are a setpoint problem
On the IEEE 39-bus case, the voltage violations of a DC OPF dispatch come from fixed generator setpoints and reactive limits, not from the real-power dispatch.
-
Do real Saildrones sail as fast as our model says?
NOAA's Saildrone surveys give a median 75 km/day. Our wing-sail model, given the same wind and course, says 2.5 times that, erasing its edge over a solar boat.
-
Does a fresher weather forecast help a wing-sail boat?
Cutting the delivery delay of NOAA HRRR wind forecasts from 3 hours to 1 hour changed a replanning boat's progress by 0.2 points, not enough to matter.
-
Does a guard that skips useless audit strata still cover at 1% prevalence?
On GoEmotions fear (1.2%) and sadness (2.2%), a pool-unless-3-positives guard covers 94.8 to 95.2% with 8 to 9% narrower intervals.
-
Does a sailing drone gain by tacking less?
Three negative results from the marine-hull programme: switch margins and minimum-dwell rules do not rescue a wind-forecast tack planner.
-
Does a weather model help a wing-sail boat steer? Four negative results
Archived HRRR forecasts, with bias correction, a trend nowcast and height correction, do not help a simulated wing-sail boat beat following the wind headers.
-
Does a wind-driven survey boat still beat a solar one if every tack costs time?
On real NDBC buoy winds a wing-sail drone on fixed survey lines keeps 73 to 87% of its free-course distance, even when each tack loses up to two minutes.
-
Does tracking the delay fix audio sync when the first window is wrong?
A delay prior tracked from earlier windows does no better than free search for audio sync on ten real clips: a series is right or wrong from its first window.
-
How big a wind-vane error can a fitted bias absorb? A cliff at 1.25 to 1.5 degrees
In a Froude-scaled aircraft model, fitting one bias per sensor channel survives a 1.25 degree angle-of-attack error and fails abruptly by 1.5.
-
How far does a real sailing drone go in a day?
A sailing-drone model said 180 to 270 km a day. Saildrone's Antarctic circumnavigation averaged 112.
-
How much does tacking cost a wing-sail drone?
A simulated 4 m wing-sail drone beating 38 km loses a mean 4.5% of the trip to tack stalls if its router ignores them; planning with the cost cuts that to 1.0%.
-
How much angle-of-attack bias can UAV system identification take?
A fitted bias repairs a constant angle-of-attack error up to 1 degree in a gust-response fit for a cargo aircraft model, and breaks between 1 and 2 degrees.
-
How much angle-of-attack sensor error can a sub-scale flight-test fit take?
A constant bias on a UAV's angle-of-attack vane breaks mode identification abruptly between 1.25 and 1.5 degrees, even when the bias is fitted.
-
Open water did not rescue the weather forecast for a sailing drone
Repeating the HRRR test at 11 NDBC buoys: the model gap shrinks but a corrected forecast only matches following wind shifts.
-
A weather model's trend did not beat 30 minutes of buoy data
Adding archived HRRR's change in wind to the buoy's own recent wind gets a sailing drone to persistence level, not above it.
-
Is the weather model wrong, or is the buoy too low?
Scaling buoy winds to 10 m cut HRRR's speed bias by more than half but left its route-replanning gap against a header-follower almost unchanged.
-
One big survey boat or several small ones for the same money?
In a simulated lawnmower survey, a fleet bought with one large boat's budget finishes in about a third of the time, unless the sonar is over half the cost.
-
San Francisco build-time intervals can be narrower, with a catch
A rule that picks tighter 80% intervals for how long San Francisco permits take to complete holds in two of three years, but only one year passes both tests.
-
A narrower prediction interval for Austin build times does not hold up
A backtest rule for tighter 80% intervals on Austin build time (issue to completion) fails on coverage and width, with the yearly numbers.
-
Narrower permit-time intervals help San Francisco and Chicago, not Austin
A rule for choosing tighter 80% intervals on permit approval time works in two cities and mostly fails in Austin, with the numbers.
-
A pitot gain error is harmless to mode fits; the offset is what hurts
In a simulated output-error mode fit, a 5% gain error on airspeed changed nothing; the constant offset a scale error implies caused all the damage.