Proof of Data
App
Open the app
Encrypt, query, verify, waitlist.
Encrypt your data
Encrypt a file in this browser to the TEE server's key.
402
Query the data
Price an AI query and pay it over x402 in USDG.
Verify the server
Check its Intel TDX proof against the onchain registry.
Private AI model
Trained on your data, pays you on every call. First training at 25 text uploads.
Join the waitlist
Uploads open in waves, in sign-up order.
Read
Protocol
Documentation
Dev log
X
Docs · How it works

The private AI model

One model trained on every text upload, inside the TEE server, paying the people whose data trained it. First training as soon as 25 text uploads are in.

Qwen2.5 0.5B Instructbase, public weightsLoRA adapterr 8, weekly, CPU/armour0.002 USDG a call
The private AI model: a public base model and its adapter slot. First training as soon as 25 text uploads are in.

The model

Version 0 is a LoRA adapter (rank 8) on Qwen2.5 0.5B Instruct. It is small and trained on CPU, which is what one confidential machine can do. Its weights never leave the server, because a model can memorise rows.

Training

Once a week the server builds the training set from live text uploads (dream journals, receipt text read inside the server, voice notes transcribed inside it), drops trap rows and deleted uploads, and trains. Fewer than 25 text uploads and nothing trains. A bigger model needs a GPU server, which the model's share of revenue pays for.

Calls and revenue

/armour costs 0.002 USDG a call over x402. All of it goes to the users, pro-rata to their uploads in the training set. Until the first training, /armour answers that the model has no weights yet.