Skip to main content
Image Forensics analyzes claim images for signs of AI generation and digital manipulation so that suspicious attachments can be surfaced to a claim handler.
Image Forensics is an optional feature gated on the Fraud Detection client feature. It is enabled per client by ValueChecker. Endpoints on this page return a Status 403 when the feature is not enabled for your client.

​
How it works

Forensics runs automatically on the images attached to the appraisals of a cart (client_reference) as part of creating the claim. Each analyzed image receives two independent verdicts, each with its own confidence score (0–1):
  • AI generated — whether the image appears to be synthetically generated.
  • Digitally manipulated — whether the image appears to have been edited. When the manipulation is localized, the result also carries the bounding boxes (regions) of the detected edits, normalized to the [0, 1] range.
When a verdict is confident enough, the image is marked as flagged and surfaced through the endpoints below. The mark is removed automatically when a later analysis comes back clean, so it never outlives the evidence for it.

​
Analyze a single image

Statelessly analyze one image and get the verdicts back directly. This does not require a cart or an appraisal — pass the image as a publicly reachable URL or a base64-encoded string, and the response returns both verdicts, their confidences, and (when manipulated) the detected edit regions.

​
Cart progress

Poll the aggregate forensics state of a whole cart. This is the progress indicator only — it reports whether analysis is still running, has finished with flagged images, or is clear. The verdicts themselves are returned by the per-image results endpoints, not here.
The cart stays in_progress while its images are still being processed and only settles once every image has been analyzed — so a transition away from in_progress is the reliable “forensics done” signal. in_progress takes precedence over has_flagged_images.

​
Per-image results

Fetch the per-attachment status and verdict so you can show the outcome for each image. Request the results for a single appraisal (scoped via the appraisal_id query parameter), or fetch the whole cart at once keyed by appraisal_id. Each attachment entry carries its processing status, the two verdicts with their confidences, the detected regions (when manipulated), and a flagged value telling you whether the verdict was confident enough to mark the image. The file_name matches the one in the appraisal’s attachments array, so you can pair a verdict with the image it belongs to.
Results are stored per claimed item rather than per appraisal, so appraisals created from the same claimed item share the same attachments and therefore the same verdicts. Appraisals with no analyzed attachments are omitted from the whole-cart response (an absent key means “no results”, the same as an empty list from the per-appraisal endpoint).

​
Clear a verdict

When a claim handler has reviewed a flagged image and accepted it as safe, you can manually clear the verdict. This marks the attachment clean and drops its flagged state the same way a clean re-analysis would.
Because verdicts are stored per claimed item, clearing an attachment clears it for every appraisal created from that item.