A frame in parts
An item is not judged as one thing. It is taken apart into what can be examined, each part is examined by its own kind of check, and the results are read back together under the confidence rule.

01
Gates first
Before anything is scored, gates decide what applies: whether there is a face, whether there is speech, whether the track is dubbed, the quality of the copy, and which parts are harmless edits. A gate never scores.
02
Facts that can decide
Two gates add facts of their own: a content credential or watermark in the file, and earlier copies of the same item. Their rules are on the confidence rule page.
03
Checks, by kind of evidence
The checks that apply then run side by side. Each belongs to one kind of evidence: image detail, change over time, the audio, the voice’s consistency, or lips against speech. Independence is counted by kind, not by the number of checks.
04
Six answers
The results become the six rows of the Media screen’s list of what was checked, each found, not found, inconclusive, not applicable or still checking, with the reason when it is not decided.
Each kind of check
Checks are described here by what they do. Which model fills each place is a setting of the installation, and every one counts only for the kinds of media it has been validated on.
| Check | What it looks at | What it cannot see |
|---|---|---|
| Faces | Where the faces are in each frame, so that the face and lip-sync checks know whether they apply and which faces a finding is on. | It gives no score. Finding a face says nothing about whether the face was changed. |
| Region | Image detail within a frame, to mark the areas that were edited or generated. | Items that are mostly text, such as memes, where text covers a fifth of the frame or more by default: the region check is turned off for them. Areas lying at least half on a logo, caption, ticker, split screen or the frame’s border are left out of the verdict. A plain crop is an edit, not a finding. |
| Whole frame | Image detail across a whole image or video frame, for signs that the whole was made by a generator. For video a face-focused check of the same kind runs as well. | Sound, and change between frames. It still runs on items that are mostly text, where the region check is turned off. |
| Change over time | How video frames follow one another, for spliced or generated stretches. | Still images, and the audio track. |
| Voice | The audio track, for a synthetic or cloned voice in all or part of it. | Under three seconds of speech, or music over the speech, by default: the check does not apply. Two voice checks count as one kind of evidence, so a voice alone can reach likely at most. |
| Speaker consistency | Whether the voice stays the same speaker across the track, for a part that was replaced. | Whether a voice is synthetic. It serves the question of a replaced part only and never counts as a voice check. |
| Lip-sync | Whether the lips move with the speech. | Items with no face or no speech, and dubbed tracks, on which it is turned off. |
When a check is skipped
A skipped check is recorded as not applicable, with its reason. It is never a score, and it never counts as having looked.
01
No face
The faces gate finds none: the face and lip-sync checks do not apply.
02
Silence or music
The voice activity gate finds under three seconds of voiced speech, by default: the voice, speaker and lip-sync checks do not apply. Where music is above its share of the track, only the voice check is turned off. Credentials and watermarks are still read.
03
Dubbing
The dub gate finds a publisher’s dub track, or speech in a different language from the earliest copy’s: the lip-sync check is turned off. Dubbing is legitimate and common in India.
04
Low quality
The quality gate grades the platform’s own copy low, medium or high. A low-quality item needs one more independent kind of evidence for any positive verdict, or it is inconclusive. Where only the product’s own lower-resolution download was graded, the grade says so.
05
Harmless edits
The edits gate finds logos, tickers, captions, split screens, crops and frame strips. Known logos and frame strips are masked before image detail is checked. Tickers and captions are never masked, and a ticker or caption whose words differ from the earliest copy’s is a finding of alteration.
Outside detectors
Some checks can be filled by outside detection services for images, video and audio.
An outside detector sits behind the same contract as the product’s own checks: it receives the same request, returns the same kind of answer, and is recorded with its kind of evidence, its media type, the quality and the language bucket of every result.
It counts toward a verdict only for a kind of media on which it has its own validated row in the ledger, below. Until then its result is shown as evidence and moves nothing. It is switched on per language bucket, and only after it has passed.
It receives key frames of the item, up to a cap set for each service. Every outside service is listed in the installation’s table of what leaves it, with the kinds of data each receives, and the configuration check fails on any that is not. Its calls count against a daily spending cap agreed for your installation, and stop at the cap.
The ledger of what has been validated
Every check, the product’s own or an outside one, counts toward a verdict only through its row in the validation ledger. The drawing shows the ledger’s shape, with every row not yet validated, as it stands before an installation’s baseline weeks; it holds no figures, and the three status marks beside it are a key, not a result.

1
Kind of check
One row for each check, by kind of evidence. A changed model is a new model: the row is keyed on the model’s own weights.
2
Media type
Image, video or audio.
3
Quality
Low, medium or high, as the quality gate grades it.
4
Language bucket
For audio, the language; Hindi audio is its own bucket, and regional languages follow.
5
Status
Every row starts as evidence only, with the reason, until it passes. Validated needs an approver’s sign-off. A validated row is tested again each month and on every model change; one that starts to flag too many unaltered items from the reference set is demoted and stops counting until it passes again.
6
The key
Filled: validated. Outlined: evidence only. Outlined in red: demoted.
Go through the checks and the ledger with your technical reviewers before any demonstration.
Request a demo/methodology/detectors · What is checked