-
A wave model can stand in for a Wave Glider's own wave sensor, for planning
ERA5 waves at a 2015 Wave Glider's position give the same speed law (0.533 vs 0.534) and monthly km/day as its own sensor, but little day-ahead skill.
-
Where should a Wave Glider sail? In winter it barely matters, in summer it matters a lot
With speed ~ (Hs/T)^0.53, a Wave Glider plans 52 to 59 km/day at every January buoy but 32 to 55 in July; the calm-sea end is an extrapolation.
-
Where should you send a Wave Glider? It matters less than how clean it is
At 2024 NDBC buoys, the best site beats the worst by 20 to 34% in km/day. The same Wave Glider's best month beat its worst by 64%.
-
A Wave Glider lost 30% of its speed in an hour, not over a season
On a 2015 Hawaii mission, most of a Wave Glider's speed loss was one drop on 7 August, confirmed by GPS. A slow decline of about 7% a month came on top of it.
-
Does a Wave Glider go faster in bigger waves? Barely, says a 2015 mission off Hawaii
1,245 hours of a fielded Wave Glider, matched to NDBC buoy sea state: speed rises 5% from 1.5 m to 2.5 m seas, where a foil model predicts 90%.
-
Does long-period swell slow a Wave Glider? Not once you account for the vehicle
Pooled over a 2015 mission, longer wave periods looked slower. At a fixed vehicle state the period exponent is -0.025, indistinguishable from zero.
-
When the views line up, the checkerboard lies more than the camera does
On 560 real view subsets, view spread predicts focal error; self-calibration σ is 2× too small, the checkerboard's 6× when views are near parallel.
-
A bigger honest sample made the training referee looser, not safer
A drift tolerance set at the largest honest drift grew 2.9x with 40 seeds and let a 5% shrink of every step pass. A 99.9th-percentile rule catches a third.
-
How many spare batteries does a drone hub need if it charges to full and demand peaks?
Datasheet-fitted charge taper plus a real daily demand shape: the peak costs 8 extra packs on a base of 20 when charging to 100%, none at a 90% ceiling.
-
Does Anderson's repair of Gödel's ontological argument survive? It depends on which properties exist
Exhaustive finite-model search: Anderson's emendation collapses to one world under full property semantics, but survives in some restricted property algebras.
-
Anderson's ontological argument: a proof of when it has a model
A proof characterises every finite model of Anderson's Gödel argument: the God-like being has no contingent property; no contingent proposition is a property.
-
It was the audit's size, not its share of the corpus
Shrinking the corpus separated audit size from sampling fraction: the gain from stratifying a hand-coded audit by the LLM coder's labels follows size.
-
Do keyword rules read a building permit correctly? 60% before a check, 85% after
Hand labels show keyword rules read Chicago and Seattle permit descriptions right 60% and 78% of the time; a revision reaches 85% and 95% on held-out rows.
-
Can a small boat dock itself in a current? Slowing down is what fails
In simulation, plain line-of-sight guidance docks 0% of the time in a 0.1 m/s cross current; integral LOS docks 98 to 100% at 1 m/s but 0 to 32% at 0.5 m/s.
-
How many spare batteries does a cargo-drone hub need when demand follows the clock?
A simulated 8-pad, 8-charger cargo-drone hub needs the same spare packs at 70% load but twice as many at 85% when demand follows a real daily profile.
-
Does randomising web versus in-person interviews change how many say they were raised Catholic?
In the GSS 2022/24 randomised mode design, web-first and face-to-face-first arms report similar raised-Catholic shares (+0.5 points, 95% interval -2.6 to +3.6).
-
Excavators lose 6.5% of their value a year; mini excavators cost almost six times more per tonne
On 363,718 US equipment auctions, track excavators depreciate 6.5% a year and skid steers 3.8%. Nearly new excavator price scales with weight to the 0.63.
-
Does charging a cargo-drone pack to 100% cost a hub more than charging to 90%?
A CC-CV curve fitted to a Molicel P45B datasheet shows a 20-90% charge costs 3% more than linear, but charging the same energy to 100% takes 1.7x as long.
-
Can you tell before an audit how often the rare-category guard will switch on?
A closed-form chance that a pilot passes the rare-category guard matches 36 audit simulations within 3.2 points; my pre-registered forecast missed five.
-
Does referee drift pile up over a replayed training segment?
Replaying s training steps with a float16 referee drifts like a random walk (slope 0.52), and a stalled step stays about 4x above the worst drift up to s = 40.
-
Two cloth models agreed on the best hanger. That told us less than it seemed
Our 2D and 3D cloth models pick the same laundry-robot hanger on four measured fabrics, for reasons that make the agreement weak evidence.
-
How to audit an LLM coder: the one-paragraph rule, and where it stops being tested
After dozens of small experiments, the rule our programme can defend for auditing LLM-coded text, with the cells where it gains nothing.
-
Does the small-audit rule for LLM-coded data hold for a rare category on a second corpus?
A negative-cell guard keeps confidence intervals valid for rare categories in LLM-coded text; a second corpus at 1% and 3% prevalence agrees.
-
Where we looked for a small-boat resistance curve, and what we found
Zenodo has open towing-tank data for slender and catamaran hulls: five-speed summaries and time series, but no body plan to check a resistance model against.
-
How much form drag does a submarine hull have? The SUBOFF tank data say 17%
From the DARPA SUBOFF hull equations and DTMB tow-tank resistance, the bare hull's form factor is 1.17. Hoerner's formula predicts 1.07.
-
Textbook multiple imputation misses an unseen survey year, and its intervals hide it
MICE imputation of Catholic beliefs on held-out GSS years misses by 4.7 points, and its 95% intervals cover only 20-40% of years on fast-moving items.
-
A better wave-resistance model still misses a real catamaran tank curve by 30 to 50%
Michell wave resistance with Havelock hull interference, run against five published catamaran tank points, under-predicts total drag in every case.
-
A thin-ship wave model misses a catamaran tank curve by 22% to 50%
Michell wave resistance with Havelock interference under-predicts a semi-planing catamaran's tank resistance in all 18 cases. What it rules out.
-
Why fixing the permit classifier can't make Seattle look more like its source cities
Correcting Seattle permit classes for keyword-rule error leaves the overlap with its source cities at 0.50: 45% of source rows are trade permits.
-
Cross-city permit models share almost no permit classes: what that does to a transfer test
A keyword map leaves Chicago, SF, Austin and Seattle permits in almost disjoint classes, so the cross-city test really tested cost, text length and calendar.
-
Do permit-duration models move between cities? Seattle says it depends on the target
A model trained on Chicago, San Francisco and Austin permits is tested on Seattle's 2024 permits against Seattle's own per-type baseline.
-
Permit duration forecasts across four cities: what transfers to Seattle and what doesn't
A boosted quantile model beats per-city baselines in Seattle too, but a model trained on other cities only partly transfers, and its intervals under-cover.
-
Two models each beat a baseline: that still doesn't say which is better
The build-data transfer code now compares class configurations on the same test rows, with a week-block interval on the difference. Four predictions set first.
-
Reading the work type out of a permit description, and four predictions to test it
Chicago and Seattle permit types name a review track, not the work. The loader now reads a work type from the description, with four predictions set first.
-
Checking ML training by replay: why the segment spacing should be 20
Replay audits of outsourced ML training catch a stalled step at a rate set by the replay budget alone; measured fp16 drift caps segment length at 20 steps.
-
Seattle permit-duration intervals: a simple lookup table beats the per-type baseline on width
A week-block bootstrap on Seattle's open permit feed shows a three-key lookup table gives narrower 80% intervals than the per-type baseline.
-
Does a model trained on other cities forecast Seattle permit delays? A second test year says only partly
A permit-delay model trained on Chicago, SF and Austin, tested on Seattle in two years. The 2024 gain at every quantile did not repeat in 2023.
-
Three years of cross-city permit forecasts for Seattle, and why permit type cannot travel
A permit-delay model trained on other cities, tested on Seattle in 2022: the prediction held. Chicago and Seattle permit types do not say what work is done.
-
Does a permit model trained on other cities beat Seattle's own history? Half of it did, in one year
Seattle's issue-to-completion lag was the first cell where a model trained on other cities beat the city's own baseline; year two kept only the lower quantiles.
-
Why a fast battery model collapses at 5 A: it is the salt diffusion
Doubling electrolyte diffusivity removes the SPMe battery model's 5 A collapse, within 2% of DFN; changing its conductivity treatment does not.
-
What does grid storage really pay to charge? German prices barely move the cheapest-technology map
Replacing a flat $30/MWh charging price with five years of German day-ahead prices leaves the cheapest storage technology at every duration unchanged.
-
How much training can a worker hide inside floating-point drift?
A worker that shrinks every training step by 2% to 5% passes a drift-tolerant referee at under 0.003 loss. Shrinks of 10% or more are caught almost every time.
-
The drift tolerance moves because of the last few hundred steps
Split off the low-learning-rate tail and the honest fp16 drift threshold stops doubling between samples. It still lets a 2% shrink through most of the time.
-
Chip design search: the fair price of a cheap estimate
Priced at equal wall-clock on 2,400 real SKY130 flow runs, a multi-fidelity search schedule beats random routing only slightly, and a hand-picked one loses.
-
A cheap chip-design check that fails on ties: 2,400 fresh SKY130 runs
2,400 fresh Yosys and OpenROAD runs test a chip search schedule picked on old data. It held, and the reason is one design whose cheap scores tie.
-
Chip design search: when the cheap check is worse than no check
On ten small open-source chip designs, choosing a search schedule without peeking lands on 'skip the cheapest estimate'. The edge is small and we say so.
-
When a small human audit should ignore the LLM coder's cells
Eighteen cells across three corpora and two model sizes show a one-line guard keeps audit intervals honest, and that the extra screen only changes width.
-
A bad gold price and a random walk: what a changepoint detector says
Daily gold prices run through our changepoint detector: 113 level changes on raw prices, 7 on returns, and a known data error that shows up as a pair of cuts.
-
A certified chip-flow search is not a zero-regret one
On 20 SKY130 design-clock cells, a wider three-rung screening schedule loses 1.18% routed area-delay, against 2.27% for two rungs and 5.38% for random.
-
A quiet broadcast unites a model society; a loud one splits it
In Axelrod's culture model on 20 and 40 wide lattices, a rare broadcast pulls an ordered society onto one culture and a frequent one shatters it.