Semantic Graph Engineering

1. Purpose of the Specification

This specification defines the technical rules, data models, and validation criteria for the construction and operation of the Semantic Graph Layer in Aivis-OS.

The goal is to model meaning, relations, and validity in such a way that:

• internal contradictions can be stored
• external interfaces exclusively expose deterministic states
• conflicts can be explicitly identified and resolved

2. Basic Rules

2.1 Determinism at the Interface (mandatory)

External outputs (Machine Interface Layer, JSON-LD, TSCL, APIs)
must not expose contradictory assertions at any time.

The following applies to every context:

• a maximum of one canonical assertion
• all other competing assertions remain internal

2.2 Multiplicity in the Graph Core (mandatory)

The internal Semantic Graph must be able to store competing assertions in parallel.

The loss of competing information is considered an architectural error.

3. Data Model

3.1 Entity (Node)

Each entity in the graph references exactly one entity from the Cluster-Level Entity Inventory.

Properties:

• entity_id (Foreign Key, stable)
• schema_type
• canonical_name

The Semantic Graph generates no own entities.

3.2 Assertion (Edge)

The Semantic Graph manages meaning exclusively in the form of assertions. An assertion is a directed, typed statement:

Assertions are First-Class Objects.

3.3 Mandatory Attributes of Each Assertion

Each assertion must have the following attributes:

4. Typing of Relations (Predicate Classes)

Untyped relations are not permitted.

Each assertion must be assigned to one of the following relation types:

4.1 Referential Relations

(e.g. mentions, about)

• weak semantic binding
• no transfer of authority

4.2 Structural Relations

(e.g. hasPart, isPartOf, subsidiaryOf)

• define hierarchies
• must not create cycles

4.3 Attributional Relations

(e.g. author, manufacturer, ceo, copyrightHolder)

• highly critical
• may only have one canonical assertion per context

4.4 Contextual Relations

(e.g. competitor, isSimilarTo, regulatedBy)

• serve for thematic classification
• capable of conflict

5. Conflict Model

5.1 Conflict Detection

A conflict exists if:

• two ACTIVE Assertions
• with identical subject + predicate + scope
• have different object values

In this case:

• Status = CONFLICT
• no automatic exposure

5.2 Conflict Resolution (Resolution Rules)

The selection of the canonical assertion is based on the following prioritization:

1. Scope Match (exact context beats generic)
2. Temporal Validity (currently valid beats historical)
3. Confidence Score
4. Governance Override (manual definition)

Only the resulting assertion receives the status CANONICAL.

6. Exposition (Collapsed State)

The Semantic Graph Layer exclusively delivers the following to the outside:

• Assertions with status CANONICAL
• exactly one assertion per (entity, predicate, scope)

All other assertions remain:

• internally retrievable
• auditable
• versionable

7. Versioning & Mutation

7.1 Changes

A change in meaning does not occur by overwriting, but by:

• Adding a new assertion
• temporal or status-related deactivation of the old one

Historical truths are preserved.

7.2 Deletion

Physical deletion of assertions is not permitted, except in the case of demonstrably faulty ingestion.

8. Validation & Testing

8.1 Graph Integrity Checks

Mandatory checks:

• Orphan Assertion Check (each assertion references existing entities)
• Cycle Detection (no circular Structural Relations)
• Determinism Check (no contradictory assertions in the Exposed Graph)

8.2 Coverage Metrics

Recommended metrics:

• Assertion Density per Entity
• Ratio of typed assertions to pure text
• Proportion of conflicting assertions (monitoring, no error)

9. Interaction with Other Layers

• Entity Inventory provides stable identities
• Semantic Graph Layer models meaning & conflict
• Transport-Safe Content Layer mirrors canonical assertions
• Projection Layer serializes deterministic states