The Armour
One model trained on every text plate, inside the enclave, owned by its uploaders. It trains when the enclave is live.
The model
The Armour v0 is a LoRA adapter (rank 8) on Qwen2.5 0.5B Instruct. It is small and trained on CPU, which is what one confidential VM can do. Its weights never leave the enclave, because a model can memorise rows.
Training
Once a week the enclave builds the training set from live text plates (dream journals, receipt text read inside the enclave, voice notes transcribed inside it), drops canary rows and deleted plates, and trains. Fewer than 25 text plates and nothing trains. A larger Armour needs a GPU enclave, which the Armour's share of revenue is for.
Calls and revenue
/armour costs 0.002 USDG a call over x402. All of it goes to the uploaders, pro-rata to their plates in the training set. Until the first training, /armour answers that the model has no weights yet.