Synthetic Diagnostic

Synthetic Diagnostic

The full graph software is exercised on a controlled benchmark with 60 generated incidents.

Purpose

The diagnostic tests whether the software can:

• associate heterogeneous observations;
• construct event relations;
• preserve provenance;
• retain contradictions and alternatives;
• represent sensor gaps explicitly;
• report graph-level evaluation consistently.

It is not intended to prove real-world cross-camera generalization.

Controlled conditions

The research record describes RGB/thermal/depth modalities, a missing-evidence rate of 0.12, and a contradiction rate of 0.35. Ground-truth entity correspondence, event graph structure, contradiction groups and missing-evidence annotations are generated alongside observations.

Results

Method	Entity F1	Event F1	Conflict F1	Missing F1	Graph score
IncidentGraph	0.943 ± 0.121	0.886 ± 0.149	0.933 ± 0.252	1.000 ± 0.000	0.923 ± 0.069
Appearance only	0.961 ± 0.103	0.868 ± 0.138	0.600 ± 0.494	0.400 ± 0.494	0.795 ± 0.122
Naive fusion	0.912 ± 0.154	0.822 ± 0.178	0.600 ± 0.494	0.400 ± 0.494	0.776 ± 0.132
Time only	0.677 ± 0.145	0.656 ± 0.113	0.600 ± 0.494	0.400 ± 0.494	0.726 ± 0.110

IncidentGraph has the highest diagnostic graph score on 45/60 incidents. Appearance-only has the highest on 15/60.

Interpretation

Appearance-only achieves slightly higher entity F1, while IncidentGraph scores more strongly on the complete evidentiary representation because it explicitly handles contradictions and missing evidence.

Claim boundary

Caution

Generated appearance bias, clock offsets, proposal quality and missingness cannot stand in for independent real sensor shifts. Publication-level end-to-end claims require real cross-camera identity association, temporal event linking, graph evaluation and controlled real-evidence stress tests.