How it works
From intent to proof.
An agent action has five links. Most systems cover the first two and call that authorisation. What happened afterwards then lives in a log — and a log can be rewritten.
Beschaffungs-Agent
Bürostühle bestellen · CHF 2 400
Intent
Der Agent nennt die konkrete Handlung — mit Hash über die Nutzlast.
Prüfung
Typ, Betrag, Ressource, Gegenpartei, Rechtsform, Frist.
Entscheid
Erlaubt, abgelehnt oder Freigabe nötig — mit Begründungscode.
Ausführung
Das Zielsystem handelt, gebunden an den geprüften Intent.
Bestätigung
Das Zielsystem meldet, was es tatsächlich tat. Abweichungen fallen auf.
Beweis
Signiert, hash-verkettet, ohne Datenbankzugriff nachrechenbar.
Typ im Mandat, Betrag unter der Grenze, Lieferant auf der Liste. Der Agent handelt — und das Zielsystem meldet zurück, was es tatsächlich ausgeführt hat.
The intent binds the payload
What gets checked is not «an order for 450», but this exact one. On redemption the hash is recomputed: a shopping cart swapped between check and execution shows up, even at the same amount.
The decision can be recomputed
Every denial names its code and the limit that applied at the time — observed value against bound. An auditor recomputes that without querying our database and without taking our word for it.
Three rules for escalation
The amount, the action type — and the legal form. The third cannot be waived: where the law requires a signature, no mandate saying otherwise helps.
The loop closes
The target system reports back what it actually executed. If that differs, the receipt says so. If it reports nothing at all, that is recorded too — a missing confirmation is a visible finding, not a silent nothing.
What Mandact cannot do: physically prevent an action. The service answers a question and writes a receipt. Whether a relying party respects a deny is up to it.