Knowledge & the Semantic Layer
Knowledge is the connective tissue between the data platform and the agent platform. It runs as a pipeline of increasing commitment: an ontology defines the object model; raw observations are promoted through candidates into canonical entities behind a human gate; canonical knowledge is fused into signed records; and a federated semantic manifold lets separate cells share meaning by exchanging signed chart projections, never raw vectors.
The property a buyer cares about is that every step is isolation-preserving and provenance-carrying, which is what allows an estate-level answer without an estate-level data pool, the very thing a data-residency regime forbids.
The pipeline
Figure C16: the knowledge pipeline. Observations and candidates are in-flight (cyan); a human-gated quarantine (amber) sits between candidate and canonical, so promotion is a governed act with a record, not a silent join. Canonical entities, fused records, and manifold projections are all signed (green). Across cells, the manifold exchanges signed chart projections, never raw vectors, which is what keeps the estate-level view without an estate-level data pool.
Table equivalent
| Stage | What it produces | Isolation / provenance property |
|---|---|---|
| 1 · Ontology | Object, link, and action types; SDK emitters generating OpenAPI, Python, and TypeScript clients | An ontology subset synchronises to edge devices, so capture-point extraction uses the same model |
| 2 · Observation → candidate → canonical | Raw observations promoted to candidate identities, then canonical entities | A human-gated quarantine decision; a provenance Merkle chain and continuity verification; promotion is recorded |
| 3 · Knowledge fusion | Signed, canonical knowledge records; node-level sheaf stitching reconciling what separate cells know | Bitemporal storage answers what did we believe and when did we believe it |
| 4 · Federated semantic manifold | A geometric object with an explicit metric and cited coordinates; a chart layer and a cross-cell layer with transition maps | Cells exchange signed chart projections, never raw vectors; a concept control plane is off by default |
| 5 · Roll-up | Coherence and dissonance read through an isolation-safe port | One of the axes of the Workspace Mind, rolling up to the node |
The ontology is the object model
The pipeline begins with an ontology: object types, link types, and action types (with undo and side effects), constraints, derivation rules, subsets, staged changes with approve/merge/reject, and evolution chains with verification. Its SDK emitters generate OpenAPI, Python, and TypeScript clients from the ontology, so the customer's object model becomes typed code rather than documentation. An ontology subset synchronises to edge devices, so extraction at the point of capture is guided by the same model the core uses.
Observation, candidate, canonical: behind a human gate
The surveillance layer promotes raw observations into candidate identities and then canonical entities, with evidence weighting, a bounded observation buffer, clustering and scoring, contradiction detection, entity signature packages, a provenance Merkle chain, and continuity verification. The step from candidate to canonical passes a human-gated quarantine decision, promotion is a governed act with a record, not a database join.
Fusion produces signed records
Knowledge fusion turns canonical entities into signed knowledge records through an observer lifecycle (scheduling, auto-tuning, failover, checkpoints), extraction and deduplication, contradiction resolution, evidence grading, and knowledge-pack build/import/export with signature verification. Node-level sheaf stitching reconciles what separate cells know into one node-level view, and bitemporal storage keeps both the belief and the moment it was held.
That bitemporal history is also load-bearing for decisions: the foresight composer's temporal-trajectory brake reads an entity's substrate-posture history from it to tell whether the entity is Improving, Stable, or Declining, and refuses a locally-positive action that would draw further long-horizon harm from a measurably declining one. Memory is not just recall. It is what lets the executive engine reason over time.
The manifold exchanges projections, not vectors
Federated semantics represent a workspace's meaning as a geometric object whose metric is explicit and whose coordinates are cited: a metric plane (the single source of truth for distance), a chart layer, and a cross-cell layer with transition maps, owner-cell election, and (the load-bearing detail) signed chart projections exchanged between cells, never raw vectors. A cross-engine query layer joins vector, graph, and relational results by canonical identity, and a concept control plane for representation steering is off by default.
That coherence and dissonance, read through an isolation-safe port, become one of the three axes of the Workspace Mind, which rolls up to the node, an estate-level view assembled from signed projections rather than a shared vector store.
Run it at the edge
An Edge Agent holds a delegated leaf of the semantic manifold and can extract typed structure from documents, both on-device. Zero-shot classification is model-free (embedding cosine over candidate labels), so it runs with no served model at all:
# Model-free zero-shot classification into candidate labels
s9 federated-semantics classify --input "cannot log in" --label billing,bug,praise
# Score an observation against an admitted chart projection (off-projection distance)
s9 federated-semantics score --store ./proj.redb --subject-area support --input "disk full"
# Extract text, then ontology-guided typed entities, from a document
s9 extract intake.pdf --output json --chunk-strategy paragraph
s9 extract-entities intake.pdf --ontology-types "Patient(name,age),Device(serial,model)"
score reads an admitted projection from a local store; extract is pure text
extraction (no model). extract-entities uses ontology-guided prompts and needs a
served on-device model. See the
local intelligence reference
for every flag.
Related topics
- Model Serving & Sovereign AI, the local embedding runtime that keeps fusion sovereign.
- Data Platform → Governance, the classification the pipeline carries end to end.
- Security → Cryptography, the signing behind knowledge records and chart projections.
- Platform Architecture → Cell Model, the cells the manifold federates across.