Trust isn’t a promise.
It’s a read-back.

Every wave of intelligent software runs into the same wall. A user gets a glimpse of how good the model is. A user feels how much friction it removes. A user does not, however, hand over the keys. Calendars, inboxes, drafts, decisions — those stay locked because the cost of being wrong is asymmetric.

The wall is not capability. The wall is trust. We’ve been calling that wall AGT — Artificial Generalized Trust — to keep the conversation honest about what we’re actually building toward.

Most products treat trust as a marketing problem. We say a product is private, secure, careful. The user reads the words and either believes them or doesn’t.

That model is wrong for AI. Trust here is a function of *read-back* — the model’s ability to surface what it has learned, in your own language, in a form you can edit or delete. The product earns the right to act by demonstrating that it understood, not by saying it will.

(a) People hand over keys in stages. Read-only first. Draft second. Send third. Schedule fourth. The product that takes the third step before earning the second collapses the trust.

(b) Trust is local. A user trusts an AI to draft a cold email and not to draft a board update. The same user. The same week. The right product respects the difference.

(c) Trust decays without read-back. If the model goes a week without showing what it has learned, the user assumes it has stopped learning.

Every Ve surface ships a read-back. The Intent Model is the most explicit one — a continuous, editable summary of what we believe about you, exposed at every level of detail. The chat surface includes its own read-back inline (every "I read X" link). The proactive surface shows its work before it asks.

That’s the technical shape of AGT in our system: read-backs at every loop. The argument of this paper is that any system without read-backs at every loop will hit the wall above and stay there.