World Graph

The third of ArcFlow's eight layers — and the layer that carries the hero. This is where typed real-world entities live: identity, topology, mutable state, mission tiers, hybrid-logical-clock provenance, and the catalog that binds graph queries to the World Store substrate beneath.

When someone asks "what is ArcFlow modeling?" the answer points here. The World Store quietly holds bytes; the World Graph is where those bytes acquire identity, become connected, and start carrying meaning. Everything above this layer — query, live, event, behavior, algorithm — exists to read and write against the typed entities the World Graph defines.

The World Graph is what makes ArcFlow a graph database rather than a column store with extra steps. It holds the adjacency lists, the label indexes, the vector indexes, the entity-resolution merges, the charting overlays, the live signals — every fact that can change after it was first written.

What lives here#

If a class is mutable — a charter correcting a play call, a telemetry correction, a derived embedding, an entity-resolution merge, a live signal — it lives in the Graph. If it is an immutable observation, it lives in the Perception Lake.

Why topology stays in the Graph#

Even when both endpoints of an edge are observation rows in the Lake, the edge itself is owned by the Graph. A (:Frame)-[:TRACKED]->(:Frame) relationship is two Lake-resident endpoints plus one Graph-resident edge. The graph engine holds the adjacency in a compressed CSR (compressed-sparse-row) layout; the Lake holds the row payloads.

This decoupling means new observations arriving in the Lake do not require Graph mutation. New edges discovered in post-processing require Graph mutation but no Lake change. Each side evolves at its own pace.

Identity is graph-resident#

Every node — Lake-resident or Graph-resident — has a stable identity owned by the Graph. The catalog resolves an identity to one of two shapes:

• Direct — (partition, row_offset) for exact-row lookup.
• Predicate — (partition, row_predicate) for property-scoped lookup such as entity_id = 'Unit-01'.

A Cypher pattern that touches a Lake-resident label compiles down to a Lake scan; one that touches a Graph-resident label runs against the in-memory tables. The agent writes the same query either way.

Why this matters for agents#

The World Graph is the durable, queryable shape an agent reasons against. It owns:

• Identity — a stable ID an agent can carry across sessions, conversations, and replay windows.
• Connectivity — the edges that turn "rows" into "a world."
• Mutable state — the only place an agent can record a correction, a merge, a derived fact.
• Indexes — the structures that make "find entities like this one" a sub-second operation.

Combined with the Perception Lake, this gives agents two surfaces with one schema: heavy columnar scans go to the Lake, low-latency graph traversal stays in the Graph, and the catalog hides the boundary at query time.

The three boundary rules#

The Lake ↔ Graph boundary is governed by three mechanical rules — see Perception Lake for the full statement. Summary:

• R1 — Identity owned by Graph.
• R2 — Mutability bright-line; Lake = immutable, Graph = mutable.
• R3 — Topology owned by Graph, exclusively.

Apply R2 first; the first rule that resolves wins.

See also#

• World Store — the durable byte substrate the Graph is a view over.
• Perception Lake — the immutable-observation sibling layer.
• Graph Model — the node / edge / property data model.
• Persistence & WAL — how Graph mutations become durable.
• Snapshot-Pinned Reads — how a reader sees a consistent point-in-time view across both layers.