Most identity systems think hard about how a name is created and not at all about what happens to it afterward. 4D-ID treats a name as something with a life: it is born, it changes, it can be corrected, and it can die. That lifecycle is what makes the name trustworthy over time, and it is what lets an AI reason about a thing across years, not just glimpse it in a frame.
A perceiving system, a camera, a robot, a drone, does not mint a new name every time it sees something. That would flood the world with duplicates. Instead a detection starts as an observation: ephemeral evidence that something is there, carried with its sensor, its confidence, and its time, but not yet a persistent identity.
An observation is not an identity. A grounding or reconciliation process first attempts to resolve it against an existing entity using identifiers, geometry, spatial context, and other evidence. An authorized naming policy may mint a new 4D-ID only when no acceptable match exists. This is how repeated sightings can remain evidence for one pallet rather than becoming three identities.
The whole point of an identity is that it survives change. A thing that moves, rotates, or is re-measured produces new state under the same name; the pose updates, the identity does not. This is the ordinary case and it is free.
Structural change is handled deliberately too. A building that gains a wing, a vehicle that is refitted, a scan frame that migrates, these are structural history events: the thing changed, but continuity is authoritative, so it keeps its identity and records what happened. Identity follows the continuity of the thing, not the constancy of its shape. A bridge is the same bridge across a decade of retrofits.
Two systems may create different identities for the same thing, and one identity may later prove to cover two things. The identity layer must preserve the authorized correction without pretending it performed the matching itself.
Two names found to mean one thing collapse into one. The earlier name wins; the other redirects to it and keeps its history. Two systems that never coordinated converge on a single identity.
One name found to cover two things becomes two, each linked back to the original. What was mistakenly one identity is corrected into the right number, with the history preserved.
This is the genuinely hard, genuinely important part, and it is where a naming layer earns its keep. Without merge and split, a shared world drifts into either duplicates or conflations. With them, the space of names converges on one name per thing over time.
See it happen: in the reconciliation demo, two upstream systems submit evidence about the same forklift. An external matcher proposes correspondence, and an authorized actor records the merge, with the loser redirecting and the survivor keeping every reference.
Loss of contact is a state of knowledge, not the end of a thing. An aircraft over the ocean, a robot in a basement, a vehicle in a tunnel, disappears and comes back. 4D-ID marks it lost-contact: its last state stays queryable with the gap honestly marked, and it returns as itself when contact resumes. Communications failure never silently kills an identity. That distinction, absent versus terminated, is exactly the kind of thing that matters for safety and accountability and that a coordinate has no way to express.
Things do end. A building is demolished, a vehicle is scrapped, a sensor is retired: the identity moves to terminal, deliberately and irreversibly, and its history remains queryable within the retention window. Terminal is reserved for real endings, destruction, retirement, or an authoritative termination, so it means something.
Virtual things need a different discipline. Anyone can create a virtual entity and everyone forgets to clean up, so virtual identities carry a lease: unless renewed, they expire on their own. Without that, a shared virtual world fills with abandoned pins and dead overlays until it is landfill. Leases keep the virtual population bounded by active interest.
An AI reasoning about the physical world is only as good as the identities it reasons over. A model that receives a fresh, unrelated detection every frame cannot form a belief about a thing; it can only describe instantaneous observations. Give it stable identities with a lifecycle, and it can do what reasoning actually requires:
Stable identity is what turns a stream of observations into a world an agent can hold in its head and act on. The lifecycle, birth, change, correction, absence, and death, is what keeps that world honest as it ages.
A coordinate tells you where something is right now. An identity, with a life, tells you it is the same something you knew before, and lets you reason about it across everything that has ever happened to it.
Watch identities handed off and acted on →
← All writing
The builder's path: the four calls, three starting points, how systems converge, and where AI plugs in. Translate once at the edge to attach a name.
Ten decisions, each against real alternatives, plus the questions we have deliberately left open.