Known limitations
What SyncLens does not do
SyncLens produces forensic evidence with a calibrated confidence band. It does not produce a verdict. These are the places where a result is weaker, and why the report is built the way it is.
Detection is probabilistic
Every result is a probability with a confidence band, not a yes or no. Run the same file twice and you get the same answer, but a small change to the media can move the result.
Read the band, not only the class. A moderate or low band means the evidence is mixed, and the report says so before it states a class.
Compression and quality weaken the signal
Heavy re-encoding, low bitrate, aggressive platform compression, and screen recordings all strip out the fine detail the model reads.
A clip that has been uploaded and downloaded several times is harder to analyse than an original file, and the confidence band will usually reflect that.
Very short and very long media
A few seconds may not carry enough signal for a confident read.
Long media is analysed in segments. A short manipulated stretch inside a long clip can still be found and marked on the timeline, but the overall class leans toward the majority of the footage. Read the per-segment view, not only the headline.
Generators the model has not seen
The model learns the traces left by the generation methods it was trained on. A newer method can leave different traces, or fewer of them.
Performance on an unseen generator is lower, and there is usually no way to tell in advance which case a given clip falls into. This is why the product is positioned as forensic analysis rather than a detector.
Benchmark scores do not transfer
Published scores are measured on curated datasets with known generators. Real-world media is messier, and a single figure hides how much the result changes from one clip to the next.
We do not publish a headline accuracy figure, because it would imply a consistency the task does not have. Every report carries a calibrated confidence band instead.
Forensic evidence, not proof
A result is one input to a decision, not the decision. It should be weighed with provenance, context, and anything else you know about the clip and its source.
For casework, the forensics workflow exists to make a report reproducible: a pinned model version, a hashed input, and exportable findings.
When only one track is usable
If a clip has no audio, or no detectable face, the report falls back to a single-track read and states that clearly.
The sync-coherence timeline is hidden in that case rather than shown empty, because a cross-track measure has nothing to compare.
When nothing is usable
If neither track can be analysed, because there is no face, no speech, or the media is corrupt, the report returns an unclassifiable state with guidance on what to try next.
It does not guess. An unclassifiable result is an honest outcome, not a failure to show.
What we will not claim
No "verified authentic". No "confirmed fake". No single number presented as a verdict.
Low confidence is stated before the result, not after it. If a report cannot support a conclusion, it says so.
These limits are the reason the product shows its work. Read the evidence and the confidence band, not only the headline result.
