-
Audit by the coder's own label: when stratifying a hand-coded sample pays
Stratifying an LLM-coded audit by corpus group buys nothing; stratifying by the coder's own label cuts the interval 15 to 25% when the coder is informative.
-
Can a 16-point pilot certify a chip-flow search? Not yet, and here is what failed
Where verified-dse stands on a pilot-checkable guarantee for chip-flow search: one pilot scheme is proved, a regret estimate from the pilot is not.
-
Gradient compression costs loss in proportion to its variance, in a small DiLoCo experiment
On a 413k-parameter GPT trained with DiLoCo, rand-k compression cost loss roughly in proportion to its variance factor omega = 1/p - 1.
-
Auditing LLM labels by the coder's own cells: no gain when the coder carries no signal
Splitting the human audit by a zero-shot Qwen coder's own label gains nothing on NaijaSenti, and a pilot floor widens intervals up to 22%.
-
Does a heel-to-toe centre of pressure help a humanoid carry faster?
A 1,440-case MuJoCo check: moving the planned centre of pressure heel to toe moves a Unitree G1 and H1 carry speed by one grid cell at most.
-
Does a load-sensing hydraulic pump pay for itself on a small excavator?
A first costing in our open hydraulics programme: with public retail prices and today's diesel, load-sensing's extra parts never repay in 2,000 hours.
-
Does an excavator's idle time decide which hydraulic circuit wins?
Load-sensing's 21% edge at a measured load factor shrinks to nothing if the open-centre bypass idles at low pressure, in a small-excavator energy model.
-
Does the cost of compressing DiLoCo updates depend on how often workers sync?
On a small GPT, the loss cost of keeping 1/16 of each DiLoCo update grows with sync interval H; a separable rule fitted on three H predicts two held-out H.
-
Does load-sensing pay on a small excavator? A first paper draft says: it depends on one number
Draft paper: load-sensing beats open-centre on a small excavator only with enough idle time, and its parts cost does not repay in fuel in 2,000 hours.
-
Does load-sensing hydraulics beat open-centre? It hangs on one unmeasured number
An open model of a small excavator's hydraulics finds the load-sensing versus open-centre ranking flips with the pressure an open-centre valve holds at neutral.
-
Does stratifying a human audit of LLM codes pay when the category is common?
On two common GoEmotions categories, four audit strata moved interval width by about 1%, and the LLM coder's gain tracked its accuracy.
-
Does top-k compression pay the same delay tax as random sparsification?
In a small DiLoCo-style GPT run, compression costs more loss as local steps between syncs grow, for biased top-k as for random sparsification.
-
Can you plan a coder-assisted audit from a pilot? Mostly not
Two repairs for a pilot-estimated audit of LLM-coded text, a per-cell floor and an independent pilot, tested on GoEmotions with a real Qwen coder.
-
Our excavator cycle had too little idle time: checking a model against a measured power trace
A public electric-excavator power trace has a load factor of 0.35; the force-model cycle we used for hydraulic circuit rankings has 0.61.
-
A gate stops one bad audio sync from ruining a camera's whole hour
Tracking a camera's audio delay from its own history locked 37% of cameras onto a wrong alias for an hour. Re-acquiring on a stronger peak cuts that to 8-10%.
-
Making a simulated coat hanger blunter stops helping at about 6 cm
Lifting the end-radius cap of a simulated 3D-designed hanger from 6 to 10 and 15 cm left the drape error flat (1.08 and 1.21 cm against 1.12 cm, 8 seeds each).
-
How far does a solar-powered survey boat go in a day, and how sure are we?
A deck-solar 5 m survey boat covers about 69 to 133 km/day depending on site; drawing power-side inputs jointly gives roughly +-10%.
-
How fast should a solar survey boat cruise? The hull you can't measure moves the answer more than the controller does
A deck-solar survey boat's best cruising speed shifts by 0.3 m/s across plausible hull constants, costing up to 9% of daily distance if you guess wrong.
-
How much energy can an excavator boom give back?
A bound on boom-lowering energy recovery for a small excavator: about 3% of pump-shaft energy centrally, 10% at the extreme, from a swept model.
-
How noisy can a flight test be before the adjustable tail stops working?
A least-squares fit of a simulated noisy free response shows what sensor noise and record length the adjustable-tail fix for a sub-scale cargo drone tolerates.
-
How noisy can the in-flight measurement be before the adjustable tail stops working?
Setting a sub-scale UAV's tail area from noisy mode estimates keeps gust response within 10% for most designs at 2% noise, but no design is reliable at 5%.
-
Idle time decides whether a load-sensing excavator circuit beats open-centre
In an open excavator-circuit simulation, load-sensing used 104% of open-centre energy with no idle time and 79 to 82% once idle matched a measured trace.
-
One number decides whether load-sensing hydraulics save fuel on a small excavator
Sourcing the load-sensing margin from a Rexroth A10VO manual reverses our earlier excavator result: the ranking turns on a margin two sources put 11 bar apart.
-
Low-communication training: compression costs loss in proportion to its noise, not its keep fraction
On a small GPT trained DiLoCo-style, unbiased compression costs loss in proportion to its measured error-to-signal ratio omega, across two compressor families.
-
Panelised vs stick framing, counted without a labour rate
What a framing crew handles on site when walls are built as shop panels instead of stick by stick, counted from the wall-framing takeoff on two houses.
-
What rough water costs a solar-powered boat, and why it should slow down
A marine-hull sweep: doubling hull resistance costs a 5 m solar boat about 21% of daily range; holding calm-water speed costs 12 to 27%.
-
When rough water comes in weeks, how much does slowing down on rough days gain a solar boat?
Simulated solar boat at five sites: slowing on rough days beats the best constant speed by 3 to 8% with a big battery, up to 15% with a small one.
-
Should you stratify the human audit of an LLM-coded rare category? Not at 1% prevalence
On GoEmotions fear (1.15% prevalence), splitting a 400-item human audit into four strata lifted nothing and cost coverage; one pooled sample covered 95.8%.
-
Should a solar survey boat slow down on rough days? A sweep, not a sea
A 5 m deck-solar survey boat simulated over ten years of sunshine at five sites: how much does backing off on rough days gain over one cruise speed?
-
Small survey boats: the hull ranking holds, the absolute range does not
Marine-hull programme: no rival hull beats a displacement monohull at equal length, but deck-solar km/day is known only to about +-10 to 20%.
-
A solar boat goes 10 to 28% further once its power figures come from datasheets, not guesses
Published hotel-load and panel figures raise a simulated deck-solar boat's range 10 to 28% at three energy-limited sites; best speed stays 1.1 to 1.5 m/s.
-
A solar survey boat's clever speed rule was mostly just a better constant speed
Held-out test on ten years of daily NASA POWER sunshine: a pack-level surge rule adds under 3% over the best constant speed for a small deck-solar boat.
-
One catalogue number flipped our excavator hydraulics ranking
Reading the open-centre neutral pressure from a real valve catalogue turned a modelled 21% load-sensing saving into an 8 to 12% loss.
-
What moves a solar survey boat's range most? Not the hull
Hotel load and panel yield move a solar survey boat's daily range 2 to 5 times more than the hull drag constants do.
-
We audited 138 digital-twin projects: only 3 name a real data source
A keyword audit of the repository's 138 twin projects finds 137 stylised, 3 with a real data source and 20 that state a paired correlation.
-
A coordinated attack breaks a median aggregator at half the coalition size
In a numpy model of 16 workers, a median stops protecting training at about 4 coordinated attackers, where independent flips needed 8. Training test pending.
-
A magnitude check caught every harmful silent bit flip in a tiny training run
In 360 injected single-bit weight flips across fp32, bf16 and fp16, rejecting any weight above 10x the last checkpoint's maximum caught 42 of 42 harmful flips.
-
A median survives 6 of 16 sabotaging workers, and diverges at 8
Real training with k of 16 workers flipping their updates: the median costs 0.014 nats at k=2, 0.14 at k=6, and diverges at k=8.
-
A proof that only half applied to our chip-search pilot
A pilot-sampling guarantee in our chip-flow search research was proved for independent pairs but used on disjoint ones. We tested whether it still holds.
-
A real LLM coder barely shrinks a rare-category audit
Qwen2.5-1.5B coded 6,683 GoEmotions comments for fear and sadness. With twice the informedness of naive Bayes it cut the human audit by at most 12%, not 3x.
-
A scale model feels the same gust load but pitches 30% harder
In a four-state model, a 1:5 Froude-scaled cargo aircraft matches full-size gust load factor within 2% but its peak pitch rate is 27 to 33% high.
-
A solar survey boat should speed up when its battery is full
Ten years of daily sunshine at five ocean sites: slowing a 5 m solar boat when low never helps, but surging when the pack is full adds 11 to 32% distance.
-
A sub-scale drone tail that only works if the airfoil surrogate is right
None of 78 tail designs that restore a 1:5 cargo UAV's flight modes stays within 5% if the real tail lift slope is 25% below the surrogate's.
-
A sub-scale cargo plane whose tail fixes its flight modes also fixes its gust response
In a Froude-scaled 1:5 cargo UAV model, 24 of 78 mode-matched tails also match full-size gust response within 5%; the similar model is 27-33% off in pitch rate.
-
A valid sequential audit that uses the coder costs more than one that ignores it
A betting confidence sequence makes a coder-assisted rare-category audit valid (99-100% coverage) but needs 3 to 14 times the oracle sample.
-
A weight bound catches bit flips, but not a cheater
A weight bound scaled by the init caught 15 of 15 harmful bit flips in a small transformer, but a value planted under it still harmed up to 60% of jobs.
-
A Wilson interval for LLM-coded audits of rare emotions
On GoEmotions fear and sadness, a pooled PPI estimate with a Wilson interval on its effective sample size covers 92 to 97% at audits of 400 to 1,000 labels.
-
A 1:5 cargo-drone model can match full-size power and flight modes, if you redesign the tail and place the CG to 6 mm
First aerolog-scaling paper draft: sections cannot fix a Froude model's drag, but a non-similar wing and tail can, with a tight build tolerance.
-
An 8 kg Froude model of a cargo UAV needs its centre of gravity within 6 mm
A 1:5 model of a 1,000 kg cargo aircraft matches full-size flight modes within 5% only if its CG is placed within +/-6 mm (aerolog-scaling, 7h).
-
Can a solar survey boat hold 1 m/s all year? NASA POWER data at five sites
A 5 m, 200 kg monohull on deck solar sustains survey speed in all 12 months near the equator but only 5 months at 60 N; annual-mean insolation hides winter.