Dustav.com

Trust & privacy

The glass box

See exactly what your agent knows and does — nothing is hidden from you.

Most AI is a black box by necessity: the "memory" is an opaque vector store, the context is a prompt you never see, and you're asked to trust that the thing running your household is what the brand says it is. Dustav inverts that. The agent is a glass box — every part of its working state is something you can open and read.

What's inside

The glass box has two halves, and together they mean nothing about what the agent knows is hidden from you.

The state you share with it — the everyday surfaces, each fully readable:

  • Calendar — every event the agent keeps for you, laid out and browsable.
  • Notes — the drafts and notes you've made together, kept at an address you own.
  • Facts — the durable household knowledge base: names, birthdays, preferences, standing schedules. This is what the agent remembers about you, in plain sight.

The Context tab — the deeper transparency. It shows exactly what the agent reads on each turn: the visible Charter (how it's meant to help), the operating rules, the live tool catalog (precisely what it can call), and the current calendar/facts state. It's measured with the same tokenizer the model uses and reconciled to a real turn — so it's not a diagram about the context, it is the context.

Put together: when you read the Context tab, you're reading the same words the model reads. There's no second, hidden version.

Why transparency is a feature, not a risk

The intuition that "showing the context makes it easy to break" is wrong for a product like this, and the reasons are worth stating plainly:

  • Trust compounds. The number-one anxiety with an AI in charge of your family's life is "what does this thing actually know and do?" A glass box answers that question by construction. You never have to take our word for it.
  • It keeps us honest. If the agent's behavior and its readable state drift apart, you'd see it. That pressure keeps the implementation clean.
  • The safety floor isn't in the box. The one thing we don't surface is the safety floor — not to hide behavior, but because it enumerates harms (self-harm, exploitation, violence) that no one should have to browse. The positive Charter is public; the hard floor governs the agent invisibly. (More on the constitution.)

Read, don't poke

The glass box is a read path. You can see everything, but you don't hand-edit the state in place. To change a fact, a note, or an event, you ask the agent — in chat — and it makes the edit coherently through its own tools. That split (the glass box reads; chat writes) is a deliberate design choice with real consequences for security and coherence; it gets its own page in read path vs write path.