Where your data lives decides everything else

Where does my data live?

In one encrypted database on the machine that runs the model. Chat, mail, money, projects and documents share the same store, and MyData opens every table, every row and a read-only SQL surface.

• A single AES-256 database per installation, not a set of cloud buckets
• MyData: table browser, read-only SQL, CSV export, every row visible
• The app publishes the inventory of every file it keeps on disk
• Backup and export are actions you run, not a vendor sync

How does it know what is sensitive?

Content is classified every time it enters a turn: yours, or someone else's. The distinction follows it into a sub-agent, a resumed run and the coordinator's result, and anything leaving the machine passes a credential and personal-identifier screen first.

• A web page, an email body or a bridged reminder list is third-party text, and is treated as such
• An instruction hidden in that text cannot originate a privileged action without a fresh approval
• Outbound tool arguments and channel messages are screened for secrets before they leave
• The classification is deterministic and on-device, not a model's guess

Which law can reach it?

The law of the place where the machine sits. No provider holds a copy to be compelled, no second country is in the processing path, and there is no region setting to trust.

• Residency is the device; sovereignty is the jurisdiction the device is in
• Localization is satisfied because the work is never sent across a border to be processed
• Inference runs locally: no provider key, no inference API, no per-token meter
• Nothing outside can suspend, geofence or rate-limit the model

Who can make it act?

A gate outside the model decides per tool and fails closed. Sending mail, moving money, writing a file and every other action that touches the outside world waits for a decision the model cannot make for itself.

• The gate decides on the tool, not on arguments the model wrote
• Autonomy rises only through recorded clean runs, and the model cannot raise its own level
• Approval can come from the desktop or from your own paired device, and it is logged
• An untrusted read forces a fresh approval on the next side effect, whatever the auto-approve setting

What record is left?

Every consequential decision goes to an append-only trail: what was asked, which model ran, which tool was called, which device approved, and when. Execution traces keep the plan, the steps and the outcome.

• The application exposes no edit or delete path for the trail
• The approval queue keeps what was approved or rejected, permanently
• A decided item is a durable record, not a transient notification
• The trail is exportable with the rest of your data