Technology and evaluation
TraceRealAI combines unfamiliar-media analysis with image watermarking. We keep the method, evaluated scope and limitations visible so each result can be interpreted responsibly.
Internal evaluation snapshot · September 2026
Current scope
Format support and evidence depth vary by media path.
| Medium | Product | Method | Boundary |
|---|---|---|---|
| Images | Detect | Face-focused or general-image analysis | A rationale is available on the general-image path; face-focused results are label-only. |
| Video | Detect | Thirty-two frames sampled across the clip | The current path does not use a temporal model, analyse audio or localise manipulation. |
| Images | Protect | CarrierSeal image watermark embed and check | Image-only. No public API, asset registry or signed credential layer is included. |
Audio-only detection and video watermarking are not currently supported.
Selected internal evaluations
These results are scoped to the named checkpoint and local test set. They are not independently audited or whole-platform guarantees.
The selected general-image checkpoint was evaluated on a balanced XPlainVerse holdout. Accuracy varied materially across generators.
Scope: one local holdout for the selected general-image checkpoint. These are not face-detector or video accuracy figures.
The selected image model was evaluated on 1,000 distinct SA-1B source images. It embeds a raw 256-bit model payload.
Scope: one local image protocol. No watermark survives every possible edit, and these figures should not be translated into a universal guarantee.
Reading results
It is a model assessment of supported visual signals, not a finding that an event is true or false.
Raw outputs and fake-frame fractions should not be read as calibrated probabilities of correctness.
A found marker means this check recovered a TraceRealAI watermark. It does not prove ownership, origin or scene truth.
Try live media detection or explore the prepared Protect workflow.