How Protection Works — Curiosity Cat

Territories and states

Every machine is a map of territories

Every agent, project and server on your machine is a territory.

GUARDED: means a protection profile is applied there and its walls have been tested — dated.
UNGUARDED: means nothing has been applied — that territory runs on whatever defaults it shipped with.
Never proved: means exactly that: we have not tested those walls, so we will not colour them green.

One level per agent, deliberately

A fleet is not uniform

Every apply is per-target. Your inbox-reading agent can be a Housecat — stays home — while a research agent runs Alley Cat — roams its own workspace and tools — and one deliberate risk-taker runs Tiger — hunts, only where you would watch it work.

A fleet is not uniform, and should not be guarded uniformly.

If one agent comes home affected

What holds, and what doesn't

Its own walls still hold — a profile does not care whether its agent has gone strange, and a compromised Housecat can barely reach anything. Other agents walls do not weaken — every profile is independent; nothing agent A does rewrites agent B's rules.

But shared spaces are corridors: agents that share folders, channels or servers can pass trouble along them. The real fleet exposure is not any one agent's level — it is the overlap of allowed territories.

Guarding agents one by one shrinks the corridors between them.

Honest scope

What this does and does not govern

Honest scope: Today's profiles govern what an agent may touch. They do not yet inspect what travels between agents — that network layer is specified, not shipped, and until it ships we say so here rather than imply otherwise.

Proof, not promises

The cat attacks its own walls

After every apply, the cat attacks its own walls. Trials that ran live are labelled observed; trials that only checked rules against themselves are labelled self-consistency — a weaker claim, labelled as one.

A dated Clean Bill, or an honest list of what did not hold.