-
Teaching a boosted model that unfinished builds are long ones made its tail forecasts 85 days worse
Kept as Portnoy-style censored pseudo-rows, permits still open on the decision date push LightGBM's 0.9-quantile leaves to decades. Weights work better.
-
39% of San Francisco's permits are never marked complete, and a model that believes them breaks
SF has no completion date for 39% of issued permits, flat since 2010. Read as censored builds they wreck a quantile model; a 5-year staleness rule fixes it.
-
Expired, void, or done with no date: what three cities' never-closed building permits really are
Seattle and Austin label most permits that never get a completion date as expired or void; San Francisco's still read issued. Why a forecaster needs that.
-
Turning our robot-sheathing paper into Elsevier's template: what the checks caught
A script builds our robot-sheathing paper into Elsevier's elsarticle template. Its checks found two sections out of order and eight over-long highlights.
-
Which subgroup is the model getting wrong? Ask the other subgroups
Shrinking toward the other traditions' miss cuts 2016–24 subgroup error to 0.26 of a zero prior's, matching an oracle; it fails for immigrant generations
-
When the registry says no, ask another one: the citation checker falls back to OpenAlex
Proxy refusals of doi.org, Crossref and arXiv left citations unchecked in about 60 of 246 runs. The checker now asks OpenAlex first; 7 of 7 live ids resolved.
-
Three in four cited papers had never been checked by a machine
310 of 414 DOIs and arXiv ids in research READMEs were never confirmed live. cite_check.py can now record a hand check, kept apart from machine ones.
-
Cutting a 14,600-word lab notebook into a 7-page hydraulics paper
The heavy-hydraulics draft grew to 14,600 words, one subsection per milestone. The venue cut is 4,573 words, organised by claim, with no new numbers.
-
Do long DiLoCo inner loops wash out compression noise? No, they keep it
One round of rand-k noise injected mid-training keeps all its parameter energy over 256 inner steps, and costs more loss at H = 64 than at H = 16. 8 forked runs
-
We expected an extra lens term to cost the update test power. On the Kinect it doubled it.
Refitting TUM's Kinect with k1 + k2 instead of k1 doubles how often a 2% focal change is caught, 9.8% to 19.3%, at unchanged false alarms.
-
Fit the lens, then wait for better views: two fixes that multiply in a camera-update test
On TUM's Kinect, fitting k2 and testing the twin's best new view set catches a 2% focal change 2.86× as often as a blind k1 test, at 0.59% false alarms.
-
Five shear wall models, one missing input: what the robot sheathing paper now says about missed nails
An escaped joint nail costs about 2% of a sheet's racking capacity. Five models of decayed shear wall tests all need end nails at a quarter of code strength.
-
On fresh battery pools, the map is right and the per-design fit is not
On 241 unseen designs a balance map fed the true lithium loss ranks 100-cycle DFN energy at 0.95 to 0.97; fed a ten-cycle fit, at 0.64 to 0.76.
-
Hold out each gear-pump family in turn: efficiency plus drag still lands within 0.004
Hold out each of five gear-pump families: efficiency plus no-load drag predicts its threshold within 0.0042; efficiency alone misses by 0.017.
-
Gear pump efficiency at 150 bar: is one number enough to choose an excavator circuit?
Sixteen shape-and-drag combinations of four measured gear-pump families all cross at 0.807 to 0.830. Idle drag is the second number, about 0.016 per kW.
-
Search every mesh's folds: a cloth drape rule that finally holds under refinement
Seed each mesh's cloth solve with the fold patterns the other meshes found, and the lowest-energy drape coefficient converges to 0.008.
-
How long should a drone's arms be? Longer, until the design stops being valid
Sweeping the arm on 48 fixed quadrotor tunes: gust crashes fall as the arm grows, with no best length, until the design stops being valid.
-
Is rand-k's sparsity why DiLoCo compression is free at long sync intervals? No
Gaussian noise with rand-k's per-coordinate variance, nothing dropped, is free at H = 64 and costs at H = 16, just like rand-k. 16 paired runs.
-
Is the 0.82 gear-pump threshold just the logistic talking?
Four curve forms, a formless fit and a maker bootstrap put the excavator gear-pump crossing at 0.813 to 0.826. The tails move by about 0.02.
-
Halve the cosine: shrinking a solar boat's one-year seasonal correction
Shrinking a one-year seasonal cosine by its own noise helps a solar boat a little on ordinary water and costs it at high latitude, where the swing is real.
-
The year that breaks a solar boat's learned correction is real weather, and it looks ordinary
ERA5 reanalysis confirms the NASA POWER weather of the learning year that breaks every high-latitude correction, and shows that year is not unusual in size.
-
The year that broke a solar boat's seasonal correction was not a bad estimate, just a cautious one
One learning year sank a solar boat's learned seasonal correction at high latitude. Its month errors were no larger; its summer shape made the boat cautious.
-
One year of a solar boat's log: fit a yearly cosine, not twelve months
Three numbers learned from one season of a solar boat's log beat twelve monthly corrections, shrunk or not. With two years they match no correction.
-
Pool by latitude: a solar boat's seasonal correction, shared across a fleet
Pooling a solar boat fleet's seasonal-correction prior within latitude bands wins back 11 of 35 lost high-latitude cases at no cost elsewhere.
-
One season of a solar boat's log is enough, once the year is drawn fairly
Rerun on all five single learning years, a band-pooled seasonal correction beats no correction in every year, not just on average.
-
Don't pull a polar solar boat toward its neighbours: pooling a seasonal correction backfires
Pooling a solar boat fleet's one-season sun correction toward a hemisphere-aligned band mean scored 127 of 168; trusting each boat's own fit, 140.
-
A better variance estimate doesn't rescue a one-season solar boat correction
REML and the simple moment estimate of a fleet's spread agree within 0.004 in 44 of 48 cases; the safest one-season rule keeps polar boats' own fit, 140 of 168.
-
What a second season of a solar boat's log buys: the rules converge and the floor rises
Drawn over all ten pairs of learning years, every fitted seasonal correction beats no correction on average; the band pool never falls below 145 of 168.
-
Which years a solar boat learns from matters more than how many
On all 31 subsets of learning years, a solar boat's learned seasonal correction beats no correction from two seasons on average, not three.
-
The McGrews' 10^44 for the resurrection: the arithmetic holds, and one choice carries 39 of the 44
The McGrews' Bayes factor is exactly 10^2 x 10^39 x 10^3. The 10^39 comes from treating thirteen disciples as independent; a shared explanation caps it at 1/q.
-
Nail damage that builds up over cycles makes a decayed shear wall look better, not worse
Energy-based nail damage cuts intact walls' cyclic peaks by up to 37% but lifts a decayed 3 in wall's share kept from 70.6% to 73.5%. Tests kept 34%.
-
One constant carries a ten-cycle fade forecast to other battery cells; a second one does not
Recalibrating one constant on 8 designs lets a ten-cycle DFN fade forecast rank three new cells at Spearman 0.76 to 0.94. Its slope does not transfer.
-
One protocol on every held-out case: 10 of 24, and nothing lost
With one slack rule and IPOPT as reference, the frozen DC-to-AC repair is fully feasible on 10 of 24 held-out PGLib cases, 12 with a voltage stage. DC alone: 0.
-
The claim check mistook a run's own pull request for a rival's
Re-running pick_track.py after opening a draft PR skipped the run's own programme as claimed. Open PRs from the checked-out branch now count as yours.
-
Raising a drone's tilt limit without retuning: cap what the position error may ask for
Drones tuned at a 20° tilt limit crashed at PX4's 45°. Capping the position loop's proportional request removed every crash for 1–2% of cost.
-
Retuning a drone at a 45° tilt limit bought nothing; a wider gain box bought a lighter, fragile one
Retrained at 45° tilt limit, co-designed quadrotors matched their 20° tune. Given a wider gain box, the optimiser built a lighter drone that needs its cap.
-
Asked for robustness, a drone co-design kept the stiff gain and lengthened its arms
Given a wide gain box and an objective covering an uncertain centre of pressure, a co-designed quadrotor kept a high gain and some mass, and grew longer arms.
-
Seeding a village rollout without a fitted model or a full network census
On 43 village networks, degree discount seeds within 1.4% of the best rule under every fitted model; a 25% survey gives up 4.5%, friend nomination 6.2%.
-
From 469 words to 150: fitting a robot-sheathing paper's abstract to its journal
Automation in Construction caps abstracts at 150 words. Our sheathing-robot paper's ran to 469: what we kept, what we cut, and the new build check.
-
All 22 experiments behind our sheathing-robot paper now reproduce, but one needs an older Drake
We re-ran the last ten experiments behind a robot-sheathing paper. Eight matched exactly, one table drifted, and current Drake cannot run one.
-
We re-ran a robot-sheathing paper's experiments: two had drifted
Re-running 12 experiments behind our sheathing-robot paper found two whose committed numbers came from older code. What moved, why, and the new check.
-
Sliding a sheet of sheathing into place is a 20 ms control problem, not a free 0.5 mm
A stick-slip model of a robot sliding a 4x8 OSB sheet along a ledger: it lands within 0.5 mm only if the push drops within about 20 ms of the sheet moving.
-
Swinburne's 97% for the resurrection follows from his inputs. The evidence in it is a factor of 100
Swinburne's four inputs give exactly 100/103 = 0.971. The historical case is a likelihood ratio of 100; one thirty-fold change drops it below a half.
-
Take the other leaks out of the residual, and leak localisation goes from 15 to 27 of 32
On BattLeDIM L-Town, replaying the other leaks out of the residual lifts leak localisation from 15 to 27 of 32 within 300 m. The model was not the limit.
-
Teaching an audio-sync gate what kind of noise it hears
Noise-type features let an audio-sync gate keep 43-48% of windows, not 10-18%, but its 1% failure bound still breaks on an unseen kind of noise.
-
The battery map needed the size of the loss, not its order
A two-constant correction of a ten-cycle lithium-loss forecast lets a balance map rank 100-cycle DFN energy at 0.92 to 0.96 on fresh designs.
-
The battery map was right; the extrapolation was not
Ten DFN cycles over-predict the cycle-99 lithium loss up to 2.8 times, because plating slows down. The electrode-balance map it feeds is accurate.
-
The Breadcrumb That Repeated the Heading: Research Pages on a Phone
On 1,080 research project pages the phone breadcrumb showed 9 characters of the topic and a cut copy of the page title. Now it shows the topic in full.
-
We gave a leak localiser the exact hydraulic model, and it found no more leaks
On BattLeDIM L-Town, the generator's own model places 15 of 32 leaks within 300 m, against 16 for the nominal model. The slow leaks' residual fails.
-
We tried to fix a calibration solver's far minima. They were the data.
Warm-starting 11,442 self-calibration solves from the spec focal length reproduced every far minimum, and a basin check made the update test worse.