A finding is only useful when you understand how it was tested and where it can fall short. Here is how we evaluate our in-house detection engine and specialist second opinions before relying on their findings.
We hold back part of every dataset for testing. We also test on images from generators a model was never trained on, plus real-world photos from different cameras and sources. This helps us see whether a model works beyond the examples it learned from.
2
Shadow mode first
Every new detector runs silently alongside live checks. We review its real-world results before switching it on for customer verdicts. Until then, its findings never change a customer verdict.
3
We look for weaknesses, not just strengths
We deliberately test hard cases: heavily compressed images, stock photography and short clips. When a model does not hold up, we keep it in evaluation rather than present it as ready.
4
Uncertain is an allowed answer
When signals disagree or are weak, we say inconclusive rather than guess. Analysts can review the context, seek a second opinion and request better source material.
5
A human makes the final decision
Scores are probabilities, not proof. Detection findings support an analyst’s judgement; a human reviews the evidence and records the final decision.
6
Licence and consent
We check the licence of every model and dataset we use on customer content; research-only data is never used for customer verdicts. Customer labels are used for training only where the organisation has opted in through its training-label setting.
7
Independent evaluation
We plan to publish benchmark results on a fixed test set, including limitations, once a detector passes real-world testing. This is a planned publication, not a claim that an independent evaluation has already been completed.
Example walkthroughs
Illustrated walkthroughs show what an analyst sees, without numerical results. These are illustrations, not live check results or customer evidence.
Example - our own test files, not customer data
Cloned voice
A synthetic text-to-speech clip is reviewed using the checks appropriate to audio. Specialist audio checks are available now; our in-house voice-clone signal remains in evaluation.
Likely synthetic speech · Needs analyst review
Upload to Analyze
The analyst uploads a synthetic text-to-speech clip as an audio file.
Audio-only checks
The clip is routed to audio-specific checks only. Visual checks do not contribute to its result.
Review likely synthetic speech
The analyst sees a likely synthetic speech finding flagged for review, listens to the clip and records the final decision.
Example - our own test files, not customer data
Sealed image
An invisible seal links a verified copy to its registered seal. The public check reveals the seal date and analyst status, not private record details.
Seal found · Seal date · Analyst status: Authentic
Seal a verified copy
After an Authentic analyst decision, the image is sealed with sealverity. The original evidence stays unchanged.
Re-save and recompress
The test copy is re-saved and recompressed. The invisible watermark is designed to survive these changes and metadata stripping; heavy edits can remove it.
Check the public stamp
On the public Check a stamp page, the seal is still found in this example. The page shows the seal date and analyst status without exposing organisation or file details.
Absence of a seal proves nothing. We do not claim that a watermark survives screenshots or cropping.
A photo contains a cloned, or copy-move, region. Edit localisation is In evaluation - rolling out: its shadow findings do not change the customer verdict.
Possible copy-move region · In evaluation
Review the edited photo
The analyst opens a test photo in which a region was copied to another part of the same image.
Inspect the heatmap
An illustrative edit-localisation heatmap outlines the cloned area and shows where to look more closely.
Keep the finding advisory
The analyst compares the marked region with the original image. The panel is labelled In evaluation and is not proof of editing or an automatic verdict.