Every AI model, at its core, is a sequence of mathematical operations. You feed in some numbers, they get multiplied, added, transformed, and eventually you get an output. A computation graph is simply a map of all those operations and how data flows between them. Catgrad, the compiler that powers Hellas, turns AI models into these graphs. This image is what one looks like after compilation. Reading left to right, three inputs are entered on the left. Each passes through a Neg box, which flips the sign of a number, positive becomes negative, and vice versa. Those results flow into Add boxes, which sum pairs together. One intermediate result gets negated again, then everything combines into a final Add on the right to produce one output. Each box is one operation. Each wire is data moving between them. This isn't a diagram someone drew on a whiteboard to explain a model. It's a rendering of the actual object that Catgrad produces, called an open hypergraph. All that term means is a structured graph defined with enough mathematical precision that there is only one correct way to evaluate it. The provider runs this object directly on their hardware. Everything is fixed before any computation happens. Every operation, every connection, every path the data takes is determined at compile time. Graphics cards normally have some freedom in how they schedule operations, which means two machines can produce slightly different results for the same computation. Catgrad's graph removes that freedom. Same graph, same input, same output, regardless of which machine runs it. This is what allows Hellas to make inference verifiable. When a job starts, both parties agree on this graph. It functions as a contract. If the output doesn't match what the graph should have produced, the client can trace through it operation by operation, find the exact box where things went wrong, and submit that single operation as proof. One operation, not the entire computation. That's what makes the verification cheap enough to actually enforce.