-
Which input decides when a hybrid cargo VTOL beats a battery one?
A global sensitivity run on a cargo VTOL sizing model: the battery's peak power, not its energy density, decides the break-even range.
-
Three permit fields carry most of a permit-duration forecast
Refitting a permit-duration model on nested field subsets: permit type, cost and work class keep 59 to 87% of the full model gain over a per-type baseline.
-
Why a bigger pilot cannot buy a guarantee on a chip-flow search
In a 40-cell chip-flow benchmark, the chance that a pilot run certifies a cost ceiling flattens at 0.45 because the cells decide it, not the pilot size.
-
Why did jittering the targets make permit-lag intervals narrower than the data?
A conformal fix for tied integer targets collapsed coverage in the same-day-heavy cells; the saved results show where, and a mechanism explains why.
-
A wind forecast with real skill still lost to a trailing mean
A ridge wind forecast fitted on real coastal data beats persistence on RMSE yet scores 0.8 points below a trailing mean when planning sailing tacks.
-
Atheism, Orthodoxy or Catholicism: the arguments machine-checked, the evidence as Bayes factors
Eleven arguments decided by a checker: the ontological argument needs system B, its atheist mirror only T. Christianity's case rests on one Bayes factor.
-
Checking the densities behind battery volume ceilings: the ranking holds
Eight reactant densities behind battery Wh/L ceilings match PubChem's handbook records; FePO4 is 6.5% off and moves LiFePO4 by 3.8%. The ranking holds.
-
A better wind forecast did not help the sailing drone tack
Coastal wind forecasts with 4 to 8% less error than persistence still did not beat a trailing-mean wind for tack planning at a 240 s tack cost.
-
Bootstrap p-values for the permit-model comparison: the same 10 of 15 cells survive
Bootstrap p-values leave the same ten significant neural-versus-boosted permit-model cells in 2024; the tiny approximate p-values were never literal.
-
A buoy's own wind history barely beats persistence
On eight NDBC buoys a fitted autoregression gains under 1% over persistence at 10 to 30 minutes and 4% at two hours; a trend rule is worse at every lead.
-
A cargo-VTOL sizing model gets the battery right and the airframe wrong
Checking a simple electric VTOL mass-budget model against two NASA concept vehicles: battery fraction matches to 0.3%, empty-weight fraction does not.
-
Process damping helps milling at low spindle speed, barely at high speed
A process-damping force raises the stable milling depth 2 to 4 times at 2000 to 3000 rpm and only 1.1 to 1.2 times above 12 000 rpm in a one-mode model.
-
Tuning a collision-avoidance planner: CMA-ES ties random search, and the gain is rule compliance
Tuning three MPC cost weights lifts held-out COLREGs compliance from about 77-84% to 86-96%; CMA-ES ties random search at 90 evaluations.
-
Do imputed religion-survey answers match the real ones?
Out-of-fold imputed Catholic GSS shares miss by a median 1.2 points across 180 decade and generation cells; cohabitation is the least calibrated item.
-
Do you need a screen on a small LLM-coded audit? A guard on three positives does most of the work
With a 1.5B LLM coder on GoEmotions, a guard of 3 pilot positives kept audit intervals valid from 1 to 23% prevalence; the extra screen mostly cost width.
-
Does an LLM-coding audit need its screen? A held-out check on GoEmotions neutral
With a weak zero-shot coder at 23% prevalence, the cell-based audit design covers about 94 to 96% and gains under 1% width over pooled. Harmless, not a win.
-
A weather forecast that only helps when you stop listening to it
At a fielded 240 s tack cost, replanning on archived HRRR wind loses to a simple header rule. A switch margin fixes the loss only by ignoring the forecast.
-
A wind forecast is worth more, not less, when a tack costs minutes
Replanning a sailing drone on a forecast beats the best rule-based tacker by 8.7 points at a 240 s tack cost, against 5.4 at 30 s. Persistence does not help.
-
Gödel's God proof only works in a world where nothing could have been otherwise
A model search shows Scott's version of Gödel's ontological argument has models only with one possible world, plus what a Bayesian ledger on God leans on.
-
How often does a 16-point pilot certify that a cheap chip-flow rung is safe? An exact formula
For verified-dse, the chance a pilot's rank-displacement certificate fires has an exact tie-aware formula that matches simulation in all 33 test cells.
-
A real weather forecast does not help a sailing drone that pays 4 minutes per tack
Replanning on archived HRRR winds lost 2 to 5 points to a header-following rule at a 120 to 360 s tack cost, though an idealised forecast gained 7 to 11.
-
Does a survey model trained on US-born Catholics work for Hispanic immigrant Catholics?
On GSS 1972-2024, a model fitted on third-generation Catholics ranks Hispanic immigrants about as well as the pooled model but misses levels by up to 13 points.
-
Does a second weather station 15 km away help predict a sailing drone's wind?
On NOAA CO-OPS 6-minute wind data, adding a station 15 to 70 km away improves a 30-minute wind forecast by under 1 point over using the target's own history.
-
NeuralFoil's lift slope is 20 to 30% too steep on a third wing section
Against measured NACA 64A010 data, NeuralFoil's lift slope is 1.23 to 1.27 times measured at Reynolds 1e5 to 2e5, as on three other sections.
-
Do neural nets really lose to boosting on permit durations? Multiple testing says: only in some years
Across 45 city, target, quantile and year cells, 32 neural-versus-boosted differences look significant; 24 survive Holm correction.
-
Does a neural net beat boosted trees at predicting permit delays? Three years, per permit
A paired bootstrap on 4 million public permits in three cities: the neural-versus-boosted difference is mostly year-specific, with one stable loss.
-
Does a neural net beat boosted trees at predicting permit timing? 45 paired tests, corrected
Bootstrap p-values for neural minus boosted quantile loss on Chicago, San Francisco and Austin permits, 2022 to 2024, with Holm and BH correction.
-
A research index that scrolled sideways on phones
The research index measured 407 px wide on a 390 px phone screen; a grid column fix brings it to 390.
-
A battery model that lets salt go negative
Two battery cell designs the fast SPMe model scores near zero show negative electrolyte concentration, while the full DFN model stays positive.
-
Stale loss factors still fix most of a DC dispatch's error
Loss penalty factors taken from the AC optimum one load step earlier keep about 90% of the gain over plain DC dispatch on the IEEE 118-bus case.
-
A sub-scale tail redesign survives a 25% lift-slope error
Cutting the tail lift slope by up to 25%, NeuralFoil's measured error, leaves the same 6 of 17 wings with a tail that restores the full-size flight modes.
-
Tuning a tack rule cannot rescue a weather forecast for a sailing drone
At a measured 240 s tack cost, archived HRRR winds lose to a simple header-following rule; a switch margin only makes HRRR copy it.
-
Seven NASA air-taxi designs say a cargo VTOL's airframe weighs 48 to 62% of takeoff weight, not 40%
A hover-plus-L/D model reproduces NASA fuel weights for two turboshaft air taxis; their non-energy weight fraction is 0.48 to 0.62, not our 0.40.
-
A pilot certificate for chip-flow search lowers regret but does not bound it
On 40 cached SKY130 place-and-route cells, a certified pilot meant zero regret in 87% of halving runs, but one uncertified-by-theory cell still lost 13.9%.
-
What a tack really costs an ocean sailing drone, and why it flips a controller ranking
Fielded Saildrone data put a tack near 218 s, not the 30 to 60 s we assumed. At that price header-following stops beating plain edge tacking.
-
Why our cargo-drone range model was 1.8x optimistic on a real aircraft
A mass-fraction range model gave 167 km for a 25 kg cargo VTOL that publishes 94 km. The gap is cruise drag, plausibly four stopped rotors.
-
Why a decent LLM coder barely shrinks the audit, and the guard that keeps it honest
A closed form caps what an LLM coder can save in a hand-coding audit, a fitted control variate captures it, and a three-positive guard keeps rare cases valid.
-
A wing lift-slope error changes which scaled wing a tail can match
Cutting the wing's lift slope by 20% in a sub-scale UAV tail search shrinks the matching wings from 6 to 5 and swaps them for different wing-loading factors.
-
How many idle rotors explain a delivery drone's range claim?
Four stopped lift rotors of the Wingcopter 198 would add enough drag to match its quoted 94 km range; all eight would overshoot it. A plausibility range.
-
Fixing one sourced number shrinks the Wingcopter drag gap
Fixing the Wingcopter 198 wingspan at its published 1.98 m shrinks the zero-lift drag needed to match its quoted 94 km range to 1.5 to 3 times a clean airframe.
-
We cut a 1.4 MB research index to 142 kB
The research index listed 1,065 project cards in one page. It now shows two per theme and fetches the rest as JSON when someone searches.
-
How long does an Austin building permit take? Real numbers from 335,000 permits
Empirical quantiles of Austin permit timing from the city's open data, 2022 to 2024, now in the build-data permit-duration prior file.
-
Neural versus boosted permit-duration models: a tie overall, slightly worse in 2024
On 45 city, quantile and year cells of open permit data, a small neural net ties LightGBM overall but trails it in 2024.
-
Neural vs boosted permit-duration models, per permit: the 2024 gap mostly doesn't repeat in 2023
A per-permit paired bootstrap on open permit data from Chicago, San Francisco and Austin finds the neural-versus-boosted difference is mostly year-specific.
-
How many checkpoints must a training verifier store?
Replaying random training steps catches a lazy worker but needs stored checkpoints. Committing fewer of them costs catch rate; here is how much.
-
Do political views explain what Catholics say about marriage? Not once you know their other answers
Fold-seed stability and a politics ablation for predicting Catholic survey answers.
-
Gradient compression in low-communication training costs nothing, then suddenly something
A small-transformer experiment finds no loss from keeping 1 in 8 coordinates of the update, but a clear loss at 1 in 16.
-
Loss penalty factors in DC optimal power flow work at full load, if you take them at the right point
On the IEEE 118-bus case, loss penalty factors evaluated at the AC optimum cut the DC dispatch's cost excess from 1.5 to 0.37 percent at full load.
-
The default audit screen covered only 85% of the time
A library check of the audit-design screen on GoEmotions found 85% coverage for a nominal 95% interval at 5% prevalence; a three-positive guard fixed it.
-
Does a pilot screen cost width when an LLM codes the text? Only when the audit is big
On 6,683 GoEmotions comments coded by a 1.5B zero-shot LLM, a pilot screen cost 17 to 29% interval width at 2,000 audited items and nothing at 500.