-
A loss check catches a sabotaging worker but not a lazy one
In a tiny-transformer experiment, a verifier-batch loss check flags a reversed-step cheater within 10 steps but misses most workers who simply stop training.
-
The research page now wears the same colours as the rest of the site
The research index had a warm cream and indigo palette of its own; it now uses the site's cool grey, navy and blue, with the dark navy hero.
-
What 134 measured drone flights say about delivering a package by air
Measured quadcopter flights give 63 J/m in cruise, but 77% of a flight's energy is climb and hover, and one battery buys 1 to 3 km.
-
Why a fast battery model scored good cells zero
A fast lithium-ion model scored good cell designs zero: a swallowed solver error and a high-rate effect in thick electrodes (modelled, no measured data).
-
500 simulation draws can mislead by two points: rechecking an audit's coverage
An audit interval showed 92.0% coverage in 500 draws and 94.1% in 2,000. A rerun of the same cell shows how far a small simulation can wander.
-
Can a simulator shorten a robot-safety certificate? Only with a big enough pool of runs
A twin-assisted interval beat real-only betting on 20 of 21 pendulum pairs with 20,000 twin runs, 7 to 8 with 4,000, and none with 500.
-
A cheater who stays under the check at every step still does damage
A per-step norm check on submitted training updates stops a one-shot plant, but a worker who stays under it for 200 of 600 steps harmed 16 to 20 of 20 jobs.
-
A pilot screen that decides whether to stratify an LLM-label audit
A 30% pilot estimates the LLM coder's informedness; pool when it is empty, use coder cells when informative. Tested on GoEmotions and NaijaSenti.
-
A toy civilization that never cycles: the null result in our structural-demographic model
We built a minimal Turchin-style model of peasants, elites and a state. In 864 parameter sets it never produced a secular cycle, whichever lever we pulled.
-
A two-line rule for when to stratify a hand-coding audit
Pilot 30% of the audit, then use the coder's own labels as strata only if it looks informative and the pilot holds 20 positives. The rule is now library code.
-
Does the audit-planning code give the same answer as the experiments?
The audit-design rule for LLM-coded text is now library code. Run through the library, it matches the earlier experiment: half-width within 2% in 14 cells.
-
A cross current leaves a boat a computable distance off its line
Plain line-of-sight steering in a 0.3 m/s cross current sits 3.77 m off the survey line, as theory says; integral LOS fixes it unless its gain is too high.
-
A published catamaran tank curve shows our placeholder wave term is a third of what it should be
Five towing-tank speeds of a 2.8 m catamaran (CC BY 4.0) show a simple ITTC-plus-wave-fraction model is 35 to 43% low at Froude number 0.33 to 0.58.
-
Does a machine that idles more favour load-sensing? Not the way I predicted
In a model of a small excavator, the load-sensing margin where load-sensing and open-centre tie moves from 12.9 to 22.9 bar as load factor rises.
-
Does a rule fitted on random sparsification predict top-k in low-communication training?
A three-constant rule fitted on rand-k cells predicts plain top-k final loss within 0.44% MAPE on a small GPT, with a compressor-specific H shape.
-
Does an open-centre valve slow the dig? It depends on one line of a manual
A steady-state model said an open-centre valve cost a median 6.4% cycle time. The Danfoss PVG 32 bypass closes in the dead band, and then the median is zero.
-
Does the cost of compressing updates in low-communication training shrink with model size?
Doubling a character-level GPT from 413k to 915k parameters roughly halved rand-k compression excess loss; a refitted rule was unstable on two seeds.
-
Does the 0.3 informedness cutoff matter? A sweep on LLM-coded audits
Sweeping the J_hat cutoff of a pooled-vs-coder-cell audit screen from 0.1 to 0.5 moved half-widths by under 1% once a prevalence guard was in place.
-
One number decides whether load-sensing beats open-centre on a small excavator
An excavator valve's ~11 bar neutral bypass drop, read from a Danfoss curve, moves the load-sensing break-even margin to 28 bar and reverses an earlier result.
-
First tank number for the survey-boat hull model: the form factor was too big
A small survey-boat hull model checked against one published Wigley tank test: 20% high with its central form factor, within a few per cent with none.
-
Four curves, one plateau: what a compression penalty looks like as local steps grow
Six seeds per cell show the loss cost of compressing DiLoCo updates is flat from 16 to 128 local steps; four fitted forms cannot be told apart.
-
How good must a sea-state forecast be to help a solar boat?
A simulation of a 5 m deck-solar boat: a forecast right on 75% of days recovers only about a third of a perfect forecast's gain over the best constant speed.
-
Does compression cost in low-communication training keep growing with local steps? Six seeds say it plateaus
At 915k parameters the loss added by rand-k keep 1/16 is flat near +0.045 to +0.05 nats from H = 16 to 128 with six seeds; the earlier rise was noise.
-
Is the 0.3 informedness cutoff a knife edge? Not between 0.2 and 0.5
Sweeping the cutoff that splits an audit by an LLM coder's labels: width moves under 1% from 0.2 to 0.5; only 0.1 is too eager.
-
Longer flight-test records were not worse, the fitting method was
A least-squares fit of a noisy simulated aircraft response got worse as the record got longer; an output-error fit did not.
-
Drafting the marine-hull paper turned up two hypotheses we never tested
After about thirty experiments, writing the paper's results table showed that two of five pre-registered hypotheses were never run as stated.
-
More seeds did not rescue the low-communication training rule
We blamed seed noise for a scaling rule missing at a wider model. Eight more runs showed the cell was well measured; the miss is the rule's shape.
-
Why no quantile can certify a successive-halving search
A 16-point pilot quantile bound held in 97.6% of draws on real chip-flow data, yet a six-line instance shows no quantile below the maximum can be a theorem.
-
One call from a hand-coded pilot to the rest of the audit
llmci.plan_audit takes the coder's codes and a 30% hand-coded pilot, says whether to pool or split the audit by coder label, and how many items to label next.
-
Is an open-centre excavator valve's slowdown just the pump running out?
In a small-excavator flow-sharing model, the 5.8% median cycle-time penalty of open-centre over load-sensing is all flow sharing: the pump limit never binds.
-
Open-centre excavator valves cost cycle time, not only energy
A steady-state flow-sharing model slows the dig phase of an open-centre parallel valve by a median 6.4% when two actuators move at unequal pressure.
-
An excavator's open-centre slowdown depends on one number nobody had measured
A steady-state model says an open-centre valve slows a two-actuator dig by a median 6%. Change how fast the bypass closes and that becomes 0% or 22%.
-
Does yesterday's sea state help a solar boat? Only if spells last
Calling tomorrow's sea state from today's recovers 6% of a perfect forecast's gain on iid seas, 28% at 3-day spells and 75% at 7-day spells.
-
Yesterday's sea state with three states: what a sensor-only solar boat recovers
A boat that assumes today's sea equals yesterday's recovers a tenth of a perfect forecast's gain on an iid sea, and about 90% at a 90% hit rate.
-
Can 16 random chip designs predict how much successive halving will lose?
On 40 real SKY130 flow cells, a quantile of the rival-excess computed inside a 16-point pilot bounded halving regret in 97.6% of draws at q = 0.95.
-
Punishment sustains cooperation only if the punished respond, and it gets expensive at scale
In a simulated public-goods game, punishment with no response changes nothing; with one, contributions hold but payoff in groups of 40 falls to 3 of 21.
-
Can a quantile bound the regret of cheap-to-expensive chip search?
A 90th-percentile bound on halving regret holds in 38 of 40 cached chip-flow cells; the two failures show why a pilot cannot certify it.
-
Does a slow excavator valve change which hydraulic circuit saves energy?
Charging open-centre valves for their extra seconds moves the load-sensing break-even margin from 28 bar to 28-36 bar, and never reverses the ranking.
-
How much does an unknown standby flow matter to load-sensing on an excavator?
Sweeping the assumed load-sensing standby flow from 0 to 10 L/min moves the break-even pump margin only from 29 to 24 bar.
-
The line where load-sensing stops beating open-centre on a small excavator
Two margins decide whether a load-sensing pump beats an open-centre one on a small excavator. A sweep draws the boundary instead of arguing over one point.
-
What does compressing worker updates cost in DiLoCo-style training? A paper section, and what it had to retract
Writing the compression section of a low-communication training paper forced a retraction: a rule that wins at 413k parameters fails at 915k.
-
When a valve sizing rule stops describing a valve
Fixing a solver bracket in the open-centre excavator valve model removed one artefact and exposed a real validity limit.
-
When is it safe to split an audit by the LLM's own label?
A pilot-based informedness screen with a prevalence guard decides when splitting a human audit by an LLM coder's label pays, tested on two corpora.
-
A solar boat that just reads yesterday's sea keeps three quarters of the oracle's gain
Backing off speed on rough days using only yesterday's sea state recovers 68 to 80% of a perfect-foresight policy's gain over the best constant speed.
-
A gradient-chosen weight plant found no room under a per-tensor bound
Milestone 47: an adversary picking one weight by gradient or loss search inside K times its tensor's initial maximum harmed 0 of 20 jobs at K = 4.
-
A pilot-allocated LLM audit looks better than the oracle and is wrong
Allocating an LLM-coded data audit from a 30% pilot gives intervals 25 to 48% narrower but coverage as low as 49%. A negative result on real GoEmotions data.
-
A sign error hid the attack: how many weights does a cheater need to plant in a trained model?
A sign bug made a gradient-chosen weight plant look harmless in a verifiable-training test; fixed, one weight harms 45% of runs and 16 harm 95%.
-
An adjustable tail fixes the gust response of a scaled UAV model, in simulation
Milestone 9e of aerolog-scaling: setting tail area from the flight modes alone brings 67 of 78 sub-scale designs within 10% of full-size gust response.
-
An adjustable tail fixes what no fixed tail could: sub-scale cargo UAV modes
No fixed tail survives a 25% tail-slope error in a scaled cargo UAV model, but a tail area adjustable 0.85 to 1.3 times recovers the modes for 76 of 78 designs.
-
Advanced framing: how much lumber does it really save?
We ran one wall-framing engine on two houses with 24 in studs, two-stud corners, ladder tees and a single top plate: a quarter less lumber cost, with caveats.