Sample reports
These examples show how different evidence patterns should be interpreted. The point is not just to display a score, but to teach teams how to read provenance, metadata, detector outputs, and visual reasoning together.
Best for showing how multiple risk signals align: no trusted provenance, stripped or weak metadata, high Sightengine-style pixel risk, a supporting Hive cross-check, and VRE findings that point to visible artifacts.
Expected handling: Manual escalation strongly recommended.
Best for explaining that a strong signed provenance chain can materially change how a reviewer interprets the file, even when detector scores are modest or incomplete.
Expected handling: Useful for lower-risk handling, but still read the full context.
Best for demonstrating why ImageVerity should not be read as a single-score oracle. If Sightengine, Hive, provenance, metadata, and VRE do not agree, the safest output is context plus escalation.
Expected handling: Do not auto-decide. Route to human review with context.
Reading rule
A useful report does not merely say "probably AI." It helps a reviewer decide whether to publish, label, hold, escalate, or investigate further. That is why ImageVerity emphasizes readable evidence chains and cautious language.
When detector signals agree, the report can recommend a stronger review action. When they disagree, the report should preserve the disagreement instead of hiding it behind a single clean-looking score.
If you want to understand why these example outcomes differ, read the methodology page and then run the workflow on one of your own images.
