Graph Reconstruction

Graph Reconstruction

Graph reconstruction converts clustered observations and event proposals into an auditable incident representation.

Reconstruction sequence

flowchart TD
    A[Observations] --> B[Cross-camera association]
    B --> C[Entity clusters]
    D[Event proposals] --> E[Map subject/object to clusters]
    C --> E
    E --> F[Event reliability score]
    F --> G[Conflict graph]
    G --> H[Selected claim + stored alternatives]
    H --> I[Duplicate suppression]
    I --> J[Typed event edges with provenance]
    K[Missing-evidence records] --> J
    J --> L[Incident timeline / graph]

Loading

Event scoring

Event confidence combines the proposal confidence with observation quality and available evidence support, with a penalty for explicit conflict in the controlled implementation.

Duplicate suppression

Local detectors may emit duplicate proposals for the same physical relation. IncidentGraph uses an event signature containing relation, fused subject, fused target and a rounded temporal bucket, retaining the strongest duplicate while preserving the retained claim's provenance.

Conflict preservation

When proposals compete for the same relation/target within the controlled temporal tolerance, the highest-scoring proposal is marked selected. Lower-scoring members are not deleted; they remain in the contradiction group as alternatives.

Output

The final representation exposes:

• fused entities;
• event relations;
• confidence;
• time;
• source evidence;
• alternatives;
• contradiction groups;
• missing-evidence objects.