Verify the pair before reading the score
A reference-based comparison should compare corresponding programme moments. If one owned candidate begins a frame later or contains a different edit, the score can reflect that mismatch as well as compression. Keep both asset identities and their preprocessing records.
Check a distinctive movement or cut near the beginning and later in the selected interval. Preserve geometry and colour assumptions too. A matching duration alone cannot establish frame-by-frame correspondence.
Do not improve an encoder preset to address a wrong reference association. First fix the comparison inputs or state clearly that the experiment measures more than encoding differences.
Use the score for its proper question
The official VMAF FAQ explains that model and viewing assumptions matter and that some artefacts fall outside its intended scope. Record the implementation, model and aggregation method rather than treating the name as one universal measurement.
Imagine a comparison whose poor section corresponds exactly to an inserted title. The next action is to inspect that editorial difference, not declare the entire codec unsuitable. After correcting the pairing, repeat only the comparison that the decision needs.
Review important passages visually at the intended presentation size. A numeric improvement cannot establish audio quality, accessibility or reliable delivery. Retain the aligned pair and the reason for any rejected run so another reviewer can distinguish a real encoding observation from a setup error.
Questions answered
Does equal duration prove frame alignment?
No. Check corresponding events and the actual timeline.
Can VMAF alone diagnose packet-delivery reliability?
Its model scope does not certify the whole delivery path.
Sources & further reading
Official references for the subjects discussed. Service terms and device software can change.