Structure first.

Most of the cost of a discovery is search, and most search is spent where no answer can be. The system works out what a problem is made of, builds a solver for exactly that, and keeps every tool it makes.

1Find the structure

Look before you search.

A hard problem is a space too large to list, with a thin region where the answers can be. Structure says where that region is: a symmetry, an algebraic identity, a way the problem splits into independent parts.

Detectors look for structures the system has met before, and it commits to one only after a checker confirms it. In blind tests it found the right structure in five of eight models, with no false detections.

A space of candidates, and the thin part of it where an answer can be.
2Compile a solver

One engine per structure.

A compiler turns the problem’s model into an engine made for it alone, with its own data layout, pruning rules and symmetry breaking. On Williamson and Legendre problems these engines run more than 30,000 times faster than Kissat and CP-SAT.

Our first engine, the FLC method, was built by hand. It runs 1,100 to 5,000 times faster than the original authors’ code on the same machine. The compiler now matches it.

Search effort as the target distance growsCompounded per-step growth. GSI engine: 4.9 to 5.4 times per extra unit of weight. Best published tool: 10 to 11 times. After six steps the gap is roughly 25 to 130 times.1×100×10k×1M×w+1+2+3+4+5+6Best published tool · 10–11× per stepGSI engine · 4.9–5.4× per step
Search effort against target distance, for our quantum-code engine and the best published tool. Six steps up, ours needs roughly 40 to 130 times less work.
3Run the loop

Optimise, simulate, experiment, verify.

Agents drive a closed loop. They choose the next bound to push, try the idea cheaply in simulation, run the compiled search at full scale, and hand the result to verification. A failed check shapes the next round.

Language models write and run the experiments. We do not rely on any one of them being right: the loop around them is what makes the output trustworthy.

The discovery loopOptimise, simulate, experiment and verify, in a cycle. Every verified result feeds a central library of detectors, checkers and solvers that the next problem reuses.LibraryOptimisepick the next bound to pushSimulatetest the idea cheaply firstExperimentrun the compiled searchVerifyre-derive it independentlydetectors, checkers, solvers
Every verified result joins the library at the centre. Every later run starts from it.
4Prove it

Nothing counts until it has been checked.

Every result we accept is re-derived by multiple separately written verifiers, so a bug in one is unlikely to repeat in the others. Where we can, it carries a certificate anyone can check without trusting us: an explicit matrix, a SAT certificate.

If a result rests on one engine, was closed only internally, may not be new, or someone else got there first, the research log says so.

5Keep everything

Each discovery makes the next one cheaper.

Every verified result leaves its detector, checker and solver in a library, and the next problem starts there. Given a kind of problem it had never seen, the system produced a verified solver in five to six minutes.

That is what we mean by generalisation: recognising known structure in a problem never seen before. It is also why the gap between one discovery and the next keeps closing.

Bring a problem with a checkable answer.

The first version opens to research teams in late November 2026. The system finds the structure, runs the search, and returns results you can verify and build on.