Research

Safety guarantees for tethered drone swarms

When several drones are tethered to one collar, can you prove they'll never tangle? I'm building a safety certificate on the braid group, the maths of how strands wind around each other, using control barrier functions to keep the swarm inside a set of safe states. So far I have working proofs for when two cables would touch, which one passes over the other, and how to track the braid as the drones move. They haven't been independently checked, and the hardest step, showing the controller can always find a safe move, is still open.

To keep the proof physically honest, I tested one of its key assumptions in simulation: what you need to know about a cable crossing to predict how it behaves. Once the cables have slid, knowing where their ends are isn't enough, but for slow motions in the simulator the position of the crossing point was. Along the way I submitted three fixes to DisMech, the rod simulator I used, and one has been merged.

The certificate also needs friction values I can trust, so I built my own frictional-contact solver. As a check, it reproduces the published capstan law for thick cables to within half a percent. With MuJoCo's default soft-contact settings and its elliptic friction cone, a block that should stay stuck creeps 13 to 26 mm over ten seconds, and a rope wrapped round a drum slips when one end pulls just 1 to 14% harder than the other, where the thick-cable law says it should hold about 5.9 times the tension. MuJoCo's no-slip option stops the creep, but its answer for the rope depends on an artificial tuning setting. Next I'm porting the solver to the GPU: the double-precision CPU version is the reference, and a batched CUDA version will be benchmarked against MuJoCo.

Fig. 6 on its own page →

Translation loss, and whether slop is load-bearing

Models now carry out instructions almost perfectly, so most failures come from getting the intent across, and whatever gets lost drifts back toward the model's own defaults. A position paper sets out that theory. The second paper asks whether the habits people call slop exist for a reason: strip out the hedges, the caveats and the flat sentence rhythm, and does the model become less honest or less safe? A first pilot ran in October.

Do coding agents understand, or recall?

Tally tests whether coding agents understand software engineering or lean on memorised patterns. It's a made-up language that looks exactly like Python but works differently in nine places. An agent gets a realistic billing codebase written in it and three tickets to close, and the grader traces each Python-style mistake back to the rule it broke. Frontier agents got through, even when they couldn't run the code; an open-weight model run locally fell back on Python habits in all three of its runs.

It's the same question as the papers above: when do models fall back on their defaults? Next I want to see whether agents keep to the spec over long tasks that get passed between sessions.

Designing a camera to spot a drone a kilometre away

How far away a camera can spot a small drone depends on how many pixels the drone covers and how faint it looks against the sky. This project asks whether a physics model, calibrated once per detector, can predict when that detector will find it on a camera it was never tuned on, so the camera can be designed on paper first. Nothing is final yet, and I rebuilt the simulator after finding errors in the first version. Along the way, on one camera none of the detectors had seen, detectors trained on simulated images found far more of the larger drones than detectors trained on real footage from a different camera.

Fig. 7 on its own page →

Devices that chirp at each other

Can a group of devices that chirp at each other work out where they are, and how the air between them is moving, without sharing a clock? Timing each chirp both ways, the trick behind two-way radio and phone ranging, cancels the clock offsets, so the layout comes out cleanly. A steady breeze looks almost exactly like the clocks being out of step, but air swirling around a loop of devices still shows, the same loop measurement ocean acousticians use for currents.

Three microphones in a triangle on one device can in principle make a steady breeze measurable too, though synchronising a few clocks does it better. So far it's theory and simulation; a bench prototype is designed but not built.

Fig. 8 on its own page →

How wind feeds waves

When wind blows over the sea, how much of its energy goes into the waves? Operational wave models answer with formulas whose coefficients are tuned against observations. I'm deriving the wind input from first principles instead, with no constants fitted to wave data, and it matches the growth rates from 25 published wind-over-wave simulation cases to within a median factor of 1.2.

A simpler model that assumes the turbulence is in equilibrium matches those simulations about as well, so they can't yet tell the two apart. On blind data from a wind-wave tank, my model predicts about 0.4 of the measured growth, and the cause of that gap is still open. Now I'm running GPU spectral-element simulations in NekRS, first reproducing a published simulation as a check, then testing at higher Reynolds numbers, where the two models' predictions differ by 20 to 55%.

Fig. 9 on its own page →

Inertial depairing: superconductors that fail after the pulse

Send a very short current pulse through a disordered superconductor, and the natural safety check is its state when the pulse ends. For pulses of about a picosecond or less in aluminium, that check comes too early. It was already known that the superconducting gap can keep falling after a fast pulse; I show that its lowest point, not the end-of-pulse state, sets how hard the pulse can safely drive, an effect I call inertial depairing. In a simplified model, a computer-checked bound puts the safe limit for a 1 ps pulse at least 20% below the end-of-pulse estimate, and numerical calculations give larger reductions. To my knowledge this correction to the threshold hasn't been calculated before. The paper proposes an aluminium experiment to test it.

Two more superconductivity studies start in early 2027: why superconductivity vanishes in overdoped cuprates, where an evidence map of about 430 published measurements is done, and whether predicted hydride superconductors survive once the pressure comes off.