Honest about what we predict, what we don't, and why our confidence intervals exist.
Every creative passes through three layers. Layer 1 (frozen): a vision + audio + language encoder predicts a per-region brain time-series. We never modify these weights. Layer 2 (learned): a calibration head per (channel × format × vertical) cell maps brain features + multimodal sidecar features to your business outcome. Layer 3 (statistical): a split-conformal wrapper produces a per-prediction confidence interval calibrated to actual outcomes in your cell.
The model gets sharper as your calibration set grows. Wider CIs at low N are honest uncertainty, not a defect.
Most neuromarketing vendors quietly include the dropped signals. We exclude them because the Boksem 2025 systematic review (n=188, k=161 trailers, 5 datasets) is the most rigorous test to date. This is the wedge — believing the replication literature.