See It Work — Sensitive data can never leak — because the road out is closed | S2 Vol 1 · Sovereign Inference & Memory

Not every job should run on the same AI. The system sorts each one into three lanes by how sensitive it is: GREEN (routine — goes to a cheap, fast AI), YELLOW (needs judgment — goes to your own high-end AI, with your approval), and RED (your most sensitive material — stays on an AI running on your own machines, never sent out over the internet).

Here's the key: RED data isn't kept safe by you remembering to be careful. The path that would send it to an outside AI is simply not built — so it can't happen, even by mistake. Once the data's in your own hands, your judgment takes over.

What this means for you

Mix sensitive and routine work together and you can't prove anything cleanly. Keep them in separate lanes — by design, not by discipline — and every routing decision becomes a fact you can audit. What this means for you: your most sensitive data can't escape to someone else's servers, because the way out was never built.

The sorting decision is itself a tracked action, so it leaves a receipt:

Routing Receipt

sensitivity labelRED

laneyour machines only

sent to an outside AI?impossible — no such path

receiptsigned & recorded

Every job you route produces this shape — which is how a year of AI work becomes an audit trail instead of a pile of leaks waiting to happen.

For the technical reader — the command, and how to verify it yourself

# one line · you do not need to run this
see walkthrough

./bl-verify
# -> the trust layer the routing receipts chain to

Full step-by-step is in Appendix RX: Hands-On Demonstrations in the book.