Actions

What it is allowed to do

An action is an object like any other. It is proposed by an actor, checked by policy, and, when it faces a customer, waits for a person. Then it happens and is recorded.

ACTIONS · TWO LINEAGES, ONE GATE The model estimates what is likely. The ontology declares what is permitted. THE PROBABILISTIC LINEAGEOntonio and the model waterfallA neural network scores what is likely from what it has seen.It is fast, fluent and sometimes wrong — so it is never the last word.◉PERCEIVE⚙DECIDE✎PROPOSEOutput: a PROPOSED fact or action, with the evidence attachedSpeech, a camera frame, a message, a document — all become proposalsTHE SYMBOLIC LINEAGEDiscern and policyThe ontology says what kinds of things exist and how they relate.Policy says who may change what. Both are data, not model output.Objectstyped at ingest, ids never changeLinksverbs between objects, declared oncePermissionswho may promote, as data the caller cannot writeOutput: KNOWN, CONFLICTED or UNKNOWN — derived at the databaseNever written by a caller, never written by a modelPOLICY + HUMANpromotes · never the modelEvery promotion into the ontology and every outbound action passes here. The model is on the left of it, always.

The same five kinds of capability that read and propose facts also bound what can act. Engines serve. Connectors, sensors and actors propose. Views read. Only a policy or a person promotes, and that includes promoting a proposed action into a real one.

Permission is data the caller cannot write. Which staff may approve which actions is a fact in the graph, evaluated by the policy engine, not a setting inside the thing asking. An automation cannot widen its own permissions.

Customer-facing outbound always queues for a human. Drafts are cheap and reversible. A sent message is neither, so the last step before a customer is a person, every time.

Worked example · A schedule change

A customer messages asking to move the appointment to two. Ontonio drafts a reply and proposes a change to the order. Policy says a same-day change needs the owner. Both sit in the queue. The owner taps approve. The reply goes out, the order updates, and both are recorded as events with provenance.

a worked example in generic terms; every operation renames the labels, the shapes do not change