lab.axe.onl · public surface
The AXE Innovation Lab is one room with two doors. On this side, our research, our model-training updates, and the live modules we use to operate CASTLE. On the other side — your portal — the same surfaces, wired to your model, your data, your evidence chain.
Training · live
Updated — · JL1 · Toronto1.19-billion-parameter foundation model · chinchilla-v3 config · trained from scratch on Canadian iron. Numbers below are the actual loss values from the live run, refreshed every minute. No simulation.
AXE trains its own foundation models on Canadian iron. Edge-1 (compact), Edge-2 (in active training above), Edge-Pro, and Edge-Flagship are one model family, four sizes, one training pipeline. Public updates land here as they pass our internal review gates.
/ now Edge-2 · 1.19B parameters · chinchilla-v3 config · 500k-step pretraining run.
That work lives one door over, behind Shield + Knox, scoped to the account that owns it.
The room-scale view of how CASTLE moves work. In the public demo it's abstract. In your portal it's wired to your model: ingestion, training, eval, audit, deployment — every tower is your tower.
Every action — ingest, train, evaluate, deploy, infer — emits a signed receipt onto a hash-linked chain. In your portal, that chain is yours. Regulator-ready, partner-shareable, your-future-self-thankable.
Dense, software-feeling, no decorative concept art. Routing, policy, retrieval, evidence, health, activity. The exact panels operators reach for when something matters.