Core service
From a shared defect list to a pack you can train on
Video-flaw packs are the main line. The other three lines fill multimodal coverage, pre-ship review, and long-tail samples when real failures are scarce.
Video-flaw packs
How the work runs
Defect types we can label (examples)
Time-axis mismatch
Objects pop in or vanish, scene layout jumps after a cut, lighting snaps with no source.
Broken motion
Jumps in movement, reversed joints, collapsed gait, hands clipping through objects.
Physics that does not hold
Wrong gravity, rigid bodies that bend, fluids that ignore conservation, collisions with no response.
Picture and identity collapse
Melted faces, character identity drift, garbled text, warped product geometry.
The other three lines
Multimodal collect and label
Image / speech / text cleanup; 2D/3D boxes, segmentation, keypoints. Specs go in the same project notes.
Eval and red team
Story continuity, brand safety, cultural sensitivity. A dossier with evidence frames — not a verbal verdict.
Synthesis and fill-in
When real samples are thin, we add marked synthetic defects and keep the source flag so they do not pollute eval sets.
What we will not do
No mystery scraped libraries. No promise to raise a closed model’s public score. No “authorized training data” made from public influencer videos.
Inquiry