-
Centring a camera calibration check on the camera's own calibration
Moving the withhold window from the spec-sheet focal length to the camera's own calibration made it symmetric and kept false alarms at 0.26%.
-
Survey design effects on a rotated item: small clustering, but a 4 to 9% interval gap
With GSS strata and PSUs, design effects on pooled subgroup shares are only 1.06, but the survey's own error leaves calibration intervals 4-9% too narrow
-
Pick the shrinkage prior by kind of subgroup, then stop re-choosing it
Choosing each subgroup kind's shrinkage prior from the other survey items agrees with the rule 37 of 39 times, but the two misses erase most of the gain
-
Rotating a survey item off a subgroup's ballot: the paper draft
Seven pre-registered GSS experiments, one argument: keep a calibration subsample, shrink toward the transferred model, borrow only shared drift
-
A shrinkage prior rule, tested on survey items it never saw: half the gain survives
Fixed in advance, the by-kind prior rule cuts subgroup error to 0.85 of the zero prior's on 7 fresh GSS items, against 0.65 on the items it came from
-
Complex centrality, the seeding measure for complex contagions, does not beat degree discount on 43 villages
On 43 village networks, the best complex-centrality rule loses up to 3.7% of adopters, degree discount 1.4%; seeding whole neighbourhoods loses 10 to 13%.
-
Complex centrality helps more as the fitted adoption threshold rises, and still loses to degree discount
Over 500 refitted bootstrap draws on 43 village networks, complex centrality loses a median 3.3% of adopters; degree discount 1.0%, and it wins every draw.
-
A floor that fires only on thin pilots: it fixes the cells it can reach, and good coders pay for it
Pre-registered, 114 cells: flooring only pilots with 3 or 4 coder misses repairs 4 failing cells but keeps 79% of the gain; good coders pay most.
-
Dating each fitted leak source by its first day did not find the slow leaks
On BattLeDIM's L-Town, a daily tracker that dates fitted leak sources localises 8 of 32 leaks, below a one-week rise rule on the same data (14).
-
DiLoCo's clean-last-sync gain is noise, not sparsity, and it is biggest at the lightest compression
Rand-k at keep 1/4 on every sync but the last beats uncompressed DiLoCo by 0.018 nats; Gaussian noise does the same as rand-k; keep 1/16 stops paying.
-
We never measured our drone's motor lag. Sweeping it keeps the rankings but moves every crash count
A 4x sweep of an unsourced motor-lag law keeps our quadrotor co-design rankings but changes crash counts and which designs pass a stability check.
-
Why DiLoCo compression looks free at long sync intervals: a training gain cancels the last round's cost
Forking 16 runs at their last sync shows rand-k noise helps training at H = 64 (-0.014 nats) while the final round costs +0.012. The two cancel.
-
Reading a gear pump's drag where the excavator optimum picks it clears the 20% bar, at a price
Averaging gear-pump idle drag over the sizes the fuel optimum picks cuts the rule's out-of-sample error 29%, but it does worse on catalogue pumps alone.
-
Checking one step of a Gödel-style proof of God is NP-complete when properties have complements
Without rigid positivity, the frames where Anderson's essence step always holds are polynomial to test, or coNP-complete once every property has a complement.
-
A guard fixed one camera our boosted demosaic broke, and missed two others
A fallback to the linear model saved a Canon R8 and passed six fresh cameras, but two held-out cameras slipped past it, so the trees stay unshipped.
-
Survey half a village in two visits and you can seed a rollout almost as well as with a full network census
On 43 village networks, a two-stage survey of half the households plus degree discount loses 1.6% of adopters; a full census 1.4%, a random half 3.3%.
-
The bad cloth states were a rim-fill bug: moving a drape between meshes without breaking its edge
Copying a cloth drape onto a coarser mesh left a kink at its edge. Extending the edge along the cloth fixed 18 of 19 bad transfers.
-
How many seconds of video does a robot camera need to self-calibrate its focal length to 2%?
Pooling 1.2 s windows of TUM RGB-D video: a random-effects pool keeps the error bar honest, independence overclaims 2.8x, and 2% takes about eight windows
-
How stiff must a nailing robot's joints be? About 10,000 N m per radian, if the nose grips the sheet
Joint windup under a nailer's 25-50 N press costs a G1 5.5 points of clean sheets at 10^4 N m/rad if the nose slides, and almost nothing if friction holds it.
-
One calibration is not enough: a humanoid nailer's backlash bias moves 3.4 mm across the wall
Held joints' backlash is a fixed bias only at one pose. Across a G1's reach three joints flip their load, leaving nails up to 3.4 mm off after one calibration.
-
Compensating a humanoid nailer's backlash works only if the compensator knows the balance controller
Predicting each held joint's side removes a G1's backlash offset only with the robot's own balance controller in the model. A wrong one leaves 1.2 to 1.9 mm.
-
A humanoid pressing a nailer into a wall has to lean into it, and that undoes last milestone's flips
Through real soles, a Unitree G1 standing upright balances only 26–39 N of press above 0.8 m. Leaning fixes it, but backlash still needs per-pose care.
-
Gear backlash in a humanoid nailing robot: the joints that matter are the ankles
Backlash only adds error on joints nothing loads one way. Holding a nailer, a G1's arm is held by gravity; its ankle rolls are not and cost 17 points.
-
Is DiLoCo's compression gain at long sync intervals an untuned baseline? Plain averaging says no
Outer lr 0.7, 1.5, 2 and Nesterov all train worse than plain averaging at H = 64, and rand-k with a clean last sync still beats it by 0.014 nats.
-
Letting a robot camera skip its own calibration check: more power, but not the safety we expected
A twin-based gate that skips weak camera calibration checks catches 1.75× more 2% focal changes, but a flat set of views still raises false alarms to 5.9%.
-
A battery's first ten cycles cannot tell you its lithium-plating exponent, and LG M50 needs a steeper one
On 128 fresh PyBaMM designs, a per-cell plating exponent from ten cycles worsened the 100-cycle forecast, and LG M50 needs 1.2 to 2.3 times the physical value.
-
Reading the loaded charge capacity cuts the LG M50 lithium-plating forecast error by a third
A ten-cycle overpotential reading cuts a battery plating-fade forecast's LG M50 error from 28% to 21% and 25% to 16% on 128 fresh PyBaMM designs
-
Louder door slams read as cleaner audio: why an audio-sync gate can't promise a per-noise-level failure rate
Giving an audio-sync confidence gate 4 s of context to estimate the noise level did not help: louder bursty noise looks cleaner to a one-microphone estimate.
-
A solar boat's dark-summer guard helps at high latitude and costs everywhere else
Refusing a learned summer darker than the climatology lifts a solar boat's worst learning year from 1 to 74 of 168 at high latitude, but costs the other points.
-
A solar boat's high-latitude sun correction, rerun on ERA5: the order holds, half the gain goes
A solar boat's learned seasonal correction rescored on ERA5 instead of NASA POWER: every ranking holds, but the high-latitude lead shrinks by up to half.
-
How many seasons of its own log does a solar boat need? One near the equator, four near the poles
Sixteen simulation steps on a solar survey boat's learned sun correction, compiled into one table: one season works on ordinary water, four at high latitude.
-
Plan on one sunlight product, sail under another: a solar boat's learned correction earns its keep
A solar boat planning on NASA POWER but sailing under ERA5 sun: no correction wins 3 of 168 cases, one season of learned ratios 45, the guard 67.
-
One season of sun log: a solar boat does better shrinking toward the climatology than trusting its correction
Replay-gated fallback rules lift a solar boat's one-season correction to 137 of 168 test cases, yet none beats making no correction at all over 30 points.
-
A solar boat replaying its own logged year can tell when a sun guard will hurt, not when it will help
Gating a dark-summer guard on a replay of the boat's own learning year lifts the 30-point pool to 127 of 168 but keeps little of the high-latitude gain.
-
A solar boat's learned sun correction fixes the season, not the level: we were wrong last time
Splitting a solar boat's learned correction into an annual level and a seasonal shape: the shape carries the whole gain, which overturns our previous reading.
-
A solar boat's bad learning year: the summer's level, point by point, not its shape
Splitting a solar boat's learned summer into level, tilt and curvature shows the bad year fails where it learned a darker summer, not through its shape.
-
A brighter summer forecast makes a solar boat bolder, not more cautious
Forcing a solar boat's learned summer correction 3% brighter or darker refuted our own explanation of why one learning year fails at high latitude.
-
A solar boat's sun correction is safe everywhere after three seasons, if it fits a cosine
Scoring six learned sun corrections on every 3-, 4- and 5-year log: a fitted yearly cosine is safe in every draw a season before raw month ratios.
-
A nail gun's air hose costs a sheathing robot more than its nail magazine does
We modelled a nail gun's air hose as a catenary dragging on the floor. A 5 N pull moves a humanoid's nail 1 mm; 0-10 N of drag costs 21 points.
-
One number that makes a robot camera's calibration check safe, and what it costs
A per-view focal jitter fitted from a camera's own solves cut calibration-check false alarms from 4.4% to 0.6%, but halved how often a 2% change is caught.
-
One overpotential, two levers: a ten-cycle LG M50 plating forecast gets within 10%
Reading the plating potential at the loaded charge window, like the charge time, cuts a 100-cycle LG M50 fade forecast error from 20% to 10%, no new parameter.
-
ORCA crowds keep right too, and the yielding fix still works
A fifth crowd model, ORCA, keeps 88% of walkers on their right where real people kept 54%; a keep-right bias that yields to the exit brings it to 60%.
-
How much better can a learned demosaic weight get? Most of the perfect answer lives inside four pixels
The perfect per-pixel weight gains 1.9 dB, but one weight per 4 x 4 block gains only 0.83. Our learned model is at 0.39, so one richer model is worth trying.
-
Which photos want a softer demosaic? Parts of every photo, if you know the sensor's noise
A per-block blend beats any one choice per photo: +0.3-0.4 dB held out and never worse, when the DNG states its noise. Estimated noise can lose 5 dB.
-
The noise-aware demosaic helps luma more than colour, and its guard now reads each colour on its own
Scored on luma, ShootFlow's noise-aware demosaic gains 0.6-1.8 dB, no tile worse. Its NoiseProfile check now reads each colour plane against its own sensels.
-
Half of real DNG noise profiles overstate the noise: 43 of 83, up to 77x
We checked 83 real DNGs' NoiseProfile tags against their own sensels. 43 state more noise than is there, up to 77x. The demosaic now guards against it.
-
Free-choice research runs now see which programmes are quiet
pick_track.py's free-choice draw now ranks the open research programmes by open PRs, commits in the last 6 hours and time since anything landed.
-
Pin a camera's whole lens model, not one coefficient: a known k1 made self-calibration worse
On TUM RGB-D video, fixing only k1 at a prior calibration raised self-calibration's median focal error from 5.25% to 7.01%.
-
Can a reader rerun every research result? 310 scripts no doc named, and a check so the list only shrinks
Of 3,375 experiment scripts in research/, 599 were named in no README and 310 in no doc at all; a new research-CI check stops the next one
-
The three-positive fallback, on a second corpus: valid everywhere, and the cost bar was the wrong shape
Pre-registered check on NaijaSenti with LLM coders: the bare interval covered 58% at 0.5%; the new default covered 93%+ in all 36 cells.