● Questions
Thirty two of them, answered plainly and grouped by what they are about.
01/08
What the two architectures are, and what laboratory validated means when we say it.
It means both architectures have been built and tested in a laboratory setting against baselines we implemented ourselves, not that they have flown. TRL 4 is where the RDI Scheme and most institutional pilots begin, and it is the honest description of where they are.
A neural architecture that carries conservation and symmetry in its construction. The invariant is a property of the object rather than a target the optimiser aims at, so it holds at machine precision for every input the model will accept, including inputs nothing in training resembled.
A neural architecture that holds knowledge locally and glues it globally. Each source carries what it can actually see, agreement is tested only where two sources overlap, and when the pieces cannot be reconciled it computes the obstruction and names the overlap where consistency breaks.
They answer different questions and are built to compose. ARCHON is for the case where a law has to hold; SHEAFON is for the case where several sources have to be reconciled and the failure has to be locatable. A problem with both properties wants both.
02/08
What the structure buys, measured against a large model, a solver, and a fitted network.
Two things. Hold a physical law exactly for every admissible input, and prove it from the trained weights without a test set. And when sources disagree, say precisely where they disagree instead of averaging them into one number. Both are available on a single machine, which is the part that changes what you can deploy and where.
A large model has to find structure in examples, and pays for that in parameters, data and compute. We put the structure in the construction, so the model stays small, trains on far less, and runs where the problem is. The guarantee reads the same on a small model as on a large one, because it comes from the construction.
Speed at the moment of decision, without giving up the physics. A high fidelity solver remains the reference and we are not trying to replace it. What we add is a model fast enough to sit inside the loop that still respects the conservation the solver respects, so the quick answer and the trustworthy answer stop being two different things.
Because the structure is the part you already know. Everything the architecture holds is something the optimiser no longer has to discover, which buys data efficiency, a smaller model, and behaviour that holds in the places your data never went.
03/08
Where this sits beside the methods a team would otherwise reach for first.
A physics-informed network puts the law in the loss function, so the optimiser trades it against fit and satisfies it approximately on the data it saw. We put the law in the architecture, so it holds for every input the model will accept, at machine precision, including inputs nothing in training resembled.
Those tell you how uncertain an answer is, calibrated on a sample. Ours tells you whether a global answer exists at all, and when it does not, which sources cannot be reconciled and where. That is a computed object rather than a confidence score.
Equivariant architectures are the closest relatives and we build against them rather than around them. The distinction is what the structure is made of: a map assembled out of quantities belonging to the object itself travels with the object, while a map fitted at one pose carries that pose with it wherever it goes. The demonstrations on the architecture pages put both to the same test, side by side.
A filter fuses sources into a single estimate and expresses their disagreement as a covariance. SHEAFON tests consistency where the sources actually overlap, and when they cannot be reconciled it names the overlap. In a degraded or contested environment the disagreement is the information, and a fused number is where that information is lost.
04/08
What holds by construction, how far it reaches, and what happens at the edge of it.
The invariant. It holds for every admissible input, by construction, at machine precision, and it can be established from the trained weights before the model is ever run. It does not depend on the data, on the size of the model, or on the input resembling anything seen in training.
It means the residual is whatever double precision arithmetic leaves behind, around one part in ten quadrillion, and not a small number that a tolerance was tuned to reach. There is no threshold to set and nothing that degrades as the input moves.
Nothing changes. The property was never inferred from a distribution, so a shift in that distribution has nothing to erode, and off distribution behaviour is the same behaviour. This is the case conventional systems handle worst, and it is the case these were built for.
It reports the disagreement and locates it, naming the overlap where consistency breaks, and withholds a global answer it cannot support rather than averaging the sources into a confident number no instrument backs.
05/08
Hardware, data, savings, and where your measurements have to live.
A desktop, a laptop or one workstation is enough to run either architecture. Everything published here was built and adversarially tested on a single machine before any external funding, which is the point of encoding the structure rather than learning it. GPU and high-performance compute raise the ceiling on the work, for finer fields, wider sweeps, ensembles and training, rather than setting the floor for a correct answer.
No. Both architectures run on a single workstation, so the computation goes to the data rather than the other way round. Nothing needs to cross your boundary.
Less than a comparable conventional model, because the part that would otherwise have to be learned from examples is already in the construction. How much less depends on the problem, and the number worth having is the one measured on yours rather than quoted from ours.
The dependency rather than the hardware. A model that holds its law by construction runs where the problem already is, so a decision does not wait on an allocation, a queue or a network hop, and the recurring cost is hardware you own rather than capacity you rent. Large machines keep their place on large work, and stop being the condition for an answer you can trust.
06/08
How the claims were tested, and what you can check without taking our word for any of it.
Yes, and that is deliberate. The guarantees are properties of the construction, so they can be checked on the trained weights rather than inferred from a benchmark, and the work runs on hardware a reviewer already owns.
The systems we compare against are implemented and measured here rather than quoted from a paper, so both sides run under the same conditions on the same hardware. A comparison against a number lifted from someone else's table is not a comparison.
You would find one admissible input where the invariant does not hold. That is the useful thing about a structural guarantee: a single case settles it, so the claim is cheap to check and impossible to dress up.
What would count as a failure is written down and timestamped before the run, so a disappointing result cannot quietly become an interesting one. A claim goes on this site only after it has survived a harness whose purpose was to break it.
07/08
The problems it is built for, and what putting it into a working system involves.
Ones where being wrong is expensive and correctness is the binding constraint. Reactor and range instrumentation, spacecraft and airframes, and any setting where several instruments have to be reconciled and a person has to sign for the result.
Fast enough to sit in the loop on the machine it was trained on, which is the property that matters for instrumentation, on board use and anything with a deadline. We measure it on your problem and your hardware, because that is the number you can act on.
Naming the quantity that has to hold, agreeing how it will be checked, and running the architecture against a baseline you already trust on data you already have. The whole of it is designed to be settled quickly and on your own ground.
Small, bounded and decisive. One quantity, one baseline, one agreed test, on data you already have. Long enough to establish the guarantee on your problem, and no longer.
08/08
Ownership, accountability, and what a first conversation looks like.
Problems where correctness is the binding constraint, and partners who want to test that claim rather than take it on trust. Write to us with the quantity that has to hold.
That is the case it is built for. A guarantee that can be checked on the weights, a failure that has a location, and a system that withholds a verdict it cannot support are all properties an accountable operator can actually use.
Agreed in writing before work starts, never assumed. What you bring stays yours, and what is built jointly is settled explicitly rather than left to be argued about later.
India. Both architectures are laboratory validated at TRL 4, built and adversarially tested on owned hardware before any external funding. Enquiries are answered by a person within two business days.
Write to us. A person reads every enquiry that arrives.
Institutional enquiries answered within two business days.