Factuam

Every AI agent you talk to today does the same thing: it reasons. It pattern matches against what it has seen and gives you an answer that sounds right. There is no experiment being run, no dataset being touched, no model being trained. It is confident guesswork and that is why it hallucinates. Factum fixes this by giving agents something they have never had: the ability to actually do the work. You give it a prompt, it finds the right dataset from Kaggle if you do not have one, and it runs a real ML experiment through a pipeline of specialist agents on the Gensyn network. Classify, plan, train, reflect, verify. Not simulated, not described, actually executed. The result comes back with a Gensyn REE proof receipt attached so you can see that the computation happened, what it produced, and that nothing was made up. The agent does not tell you what would probably work. It runs it and shows you.

The core idea is that agents should be able to do things, not just say things. The pipeline is a TypeScript orchestrator running inside a single Gensyn AXL node, where each specialist agent hands off to the next: classify the input, design the experiment, train a model, reflect on what happened, verify the output, then answer. AXL handles transport between stages using Ed25519 peer identities so every step is attributable. When a user comes in with just a prompt and no dataset, the Kaggle connector layer intercepts it before it even reaches the node, finds the most relevant dataset, and injects it as a proper payload. The actual training runs on a local Python worker that does real model fitting on that data. Once training and reflection finish, the verify stage calls Gensyn REE from inside the node to check integrity and produce a signed receipt tied to the outputs. Nothing leaves the node unverified. We kept it to a single node deliberately for the prototype to show the loop clearly: prompt in, real computation happens, verified result and proof come out.