The TheAlert manifesto

Honest label: TheAlert is a start document, in discovery. It does not exist in code, it is not a commercial offering and it has no client. This note records a position, not a measured finding, and it is the only TheAlert note until discovery produces a measurement.

What TheAlert is

Hummand’s product and the first initiative of its AI division: executable authority over the actions of AI systems, bound to a real human, with layered intervention and verifiable evidence. Every action an agent takes goes through a gateway and only happens within a limit that an identified person granted and can withdraw. An intervention order blocks new actions across the integrated environments and produces a signed record of what stopped, what is still pending and what is unknown.

The division of labour with Hummand identity is a two-column table. Identity answers who took part and which act was proven, and issues the humanproof/1 receipt. TheAlert answers which action is allowed, how to restrict it and which result was observed, and uses that receipt to bind every intervention order to a person. TheAlert produces neither receipts nor biometric results; identity does not modify the supervised infrastructure.

The manifesto

Machines will act faster than any person can follow. That is not a risk to avoid. It is the condition we design from.

Delegation is not abdication. Every power granted to a machine has a human origin, a declared limit, and a way back.

An alert is not an alarm. It is the moment human attention is called, with time and means to respond.

Proof over trust. What a third party cannot verify has not been shown.

Stating the limit is part of the product. Every guarantee carries the perimeter where it holds, and what lies outside it.

We build so that machine autonomy can grow without shrinking the human capacity to intervene.

What this text is, and what it is not

• It is a position. The sentences that any future TheAlert promise has to keep sustaining after the limits are read. If a capability contradicts one of them, the capability is what is wrong.
• It is not a result. No sentence was measured, no mechanism was reproduced by a third party. When there is a measurement, it comes the way the other notes in this house do: dataset, protocol, limitation.
• It is not an offering. There is no capability in the catalogue, no price, no deadline and no client. The start document calls this discovery: finding a concrete case, a buyer, a perimeter that can be closed and a pilot someone pays for. Only after that is there a product.
• It was published ahead of schedule. The start document planned to publish the manifesto together with the evidence format and the verifier, after the technical proof. The founder decided to publish it earlier, on 21 September 2026, and this note says so instead of hiding it.

What already exists at Hummand, and what TheAlert would add

The agents line described in the previous notes (a mandate with scope granted with live presence, a delegated act verified at the executor, a signed receipt verifiable offline, authority brought down by a human act) is the base TheAlert consumes. What TheAlert would add, and what does not yet exist in code, is the rest: an action policy enforced outside the model, the same intervention semantics across more than one environment, and a signed report that says what stopped, what is pending and what is unknown.

Two sentences from this house still hold here. A guarantee is only valid within the perimeter where it was tested. And what a third party cannot verify has not been shown.

Not to be confused with Alerts

Alerts, at hummand.com.br/en/alertas, are three situations the register of acts exposes on read: expired proof, act outside mandate, expected act missing. TheAlert is something else: supervision of the actions of AI systems. It reads those signals; it does not replace them.