video ─→ ffmpeg 1fps sample → N frames
N frames ─→ CLIP ViT-B/32 vision encoder ─→ (N, 512) stim_emb
"general audience" ─→ CLIP text → demo_emb
cortexa_v10 head per frame → per-frame curves (engagement, audience_drive, ...)
                              ↓
                    aggregate to 4 Create pillars:
                      hook            = mean engagement, first up-to-3 frames
                      hold            = mean engagement, all frames
                      algorithmic_fit = mean audience_drive
                      brand_lift      = mean demographic_appeal
                      overall (win)   = sigmoid(audience_drive_mean · 6)