Federated Semantics
Federated Semantics is how separate cells share meaning without sharing data. A workspace's semantics are represented as a geometric object whose metric is explicit and whose coordinates are cited, and cells reconcile their understanding of the same domain by exchanging signed chart projections, never raw vectors, never a shared embedding store. That single property is what lets an estate answer a question across cells without an estate-level data pool, the very thing a data-residency regime forbids.
It sits alongside the broader Knowledge & semantic layer pipeline (Federated Semantics is stage 4 of that pipeline) and is a distinct capability from the virtual graph: the virtual graph joins data across engines, while Federated Semantics reconciles meaning across cells.
The layered manifold
The semantic manifold is built in layers, each with a single job, so distance, representation, and cross-cell reconciliation never collapse into one opaque vector space.
*Figure: the federated semantic manifold. The metric plane is the single source of truth for distance; a workspace's chart layer holds cited local coordinates (in-flight, cyan); the cross-cell layer carries transition maps and owner-cell election. Cells exchange signed chart projections (green) and never raw vectors; the sheaf-coherence layer reconciles overlapping charts into a node-level roll-up. The concept control plane (amber) is off by default. Ceiling: describe the layers and the exchange model, there is no production track record for cross-cell semantic exchange to claim.
Table equivalent
| Layer | What it is | Isolation / provenance property |
|---|---|---|
| Metric plane | The single source of truth for "distance" between concepts | One authoritative metric, not a per-query similarity heuristic |
| Chart layer | A workspace's local coordinate system, with coordinates that are cited | Read scoped to the workspace that owns the chart |
| Cross-cell layer | Transition maps between charts, plus owner-cell election | Reconciliation is explicit and owned, not an implicit merge |
| Signed chart projection | The unit exchanged between cells | Signed, and it is a projection: never a raw vector or a shared embedding store |
| Sheaf-coherence layer | Reconciles overlapping charts across the chart-overlap graph | A node-level view assembled from signed projections, not a pooled store |
| Concept control plane | Glass-box representation steering | Off by default |
| Roll-up | Coherence and dissonance read through an isolation-safe port | One axis of the Workspace Mind, rolling up to the node |
Why projections and not vectors
Sharing raw vectors would rebuild the estate-level data pool that residency law forbids, the embeddings are the data. A signed chart projection carries only the reconciliation a peer cell needs to align its own coordinates, it is signed at origin so a receiver can verify where it came from, and it never reconstructs the source cell's records. This is the load-bearing detail of the whole layer: an estate-level answer, assembled from signed projections, with no estate-level data pool.
Distinct from the virtual graph
Federated Semantics is frequently confused with the virtual graph; they solve different problems and are worth stating apart:
| Federated Semantics | The virtual graph | |
|---|---|---|
| Reconciles | Meaning across cells | Data across engines |
| Unit of exchange | Signed chart projections (never raw vectors) | Query results joined by canonical identity |
| Home plane | The AI / semantic layer | The data platform |
| Scope | Cross-cell, node roll-up | Cross-engine, within reach of a query |
A cross-engine query layer does join vector, graph, and relational results by canonical identity, but that join happens within the data platform, not by pooling vectors across cells.
Where it runs across the matrix
- Single-cell: the metric plane and chart layer are fully useful on one cell; there is simply no cross-cell exchange to perform.
- Multi-cell: the cross-cell layer, owner-cell election, and signed projection exchange come into play; this is the scale-out corner the layer exists for.
- Sovereign / air-gapped: because the exchange is a signed projection, an air-gapped site can carry projections on a signed offline bundle rather than a live link, and still never ships raw vectors.
Related topics
- Knowledge & the semantic layer, the full pipeline of increasing commitment that Federated Semantics is the fourth stage of.
- The AI system (core + edge), how this layer fits the whole AI system.
- Data Platform, the virtual graph and the cross-engine query layer.
- Security → Cryptography, the signing behind chart projections.
- Platform Architecture → Cell model the cells the manifold federates across.