Evidence Model

Evidence Model

IncidentGraph uses neutral evidence records so that camera models, rules, external sensors or human annotations can enter the same reconstruction layer.

Observation record

Conceptually, an observation contains:

camera_id
modality
local_track_id
entity_type
start_time
end_time
world_position
appearance_descriptor
quality
evidence_reference

The evidence reference is immutable from the point of ingestion and is the foundation for provenance-preserving reconstruction.

Event proposal

A local recognizer or rule engine can produce an event proposal containing:

event_type
subject_observation
optional_object_observation
local_time
confidence
evidence_set
contradiction_flag

This interface decouples the graph layer from a specific activity-recognition model.

Incident graph

The reconstructed graph can be summarized as:

G = (V, E, C, M)

where:

• V — entity and location nodes;
• E — event relations;
• C — contradiction groups and alternative hypotheses;
• M — missing-evidence objects.

An event edge stores source entity, target entity, relation, confidence, timestamp, provenance and alternatives.

Evidence invariants

IncidentGraph is designed around three invariants:

1. Accepted event edges have supporting provenance.
2. Missing evidence is not converted into negative evidence.
3. Losing contradictory hypotheses remain auditable alternatives.

These invariants are schema/reasoning properties rather than properties automatically guaranteed by a generic graph neural network.