This page explains the customer-facing score definitions shown in Cortex. It intentionally avoids private implementation details.
Final score
The result page uses:
Final score = 0.65 x Brand winner-likeness + 0.35 x General attention
Brand winner-likeness is the brand-specific part. General attention is the broader attention-pull part.
This design prevents a candidate from scoring well only because it is flashy. It still needs to resemble the brand's winners.
Brand winner-likeness
Brand winner-likeness is a 0 to 100 score. Above 50 is winner-like. Below 50 is loser-like.
It can be produced from:
- Video score: the whole candidate compared with the brand signature.
- Moment score: scored creative moments compared with equivalent winner and loser moments.
- Anchor-window score: fixed timing windows such as hook, product intro, proof, offer, and CTA compared with the same windows in winners and losers.
- Creative-profile score: Gemini-derived creative psychology and structure compared with the brand's winner and loser profiles.
The active blend uses the calibrated weights for the signals available on that candidate:
Brand winner-likeness = normalized calibrated blend of video, moment, anchor-window, and creative-profile scores
When calibration has enough data, Cortex selects the mix that best separates approved winners from losers. If a signal is unavailable on a specific run, Cortex removes that component and normalizes the remaining active weights. The calibration page shows the selected weights.
General attention
General attention is derived from the candidate's attention curve. It is independent of the reference bank. The app summarizes the stronger half of the curve so a single weak patch does not erase a video that repeatedly regains attention.
Use general attention to judge attention pull. Use brand winner-likeness to judge brand fit.
Confidence
Confidence is not part of the final score formula. It is an interpretation layer.
Confidence considers:
- Winner and loser reference counts.
- Balance between the two sides.
- Scored event coverage.
- Event-type depth in the signature.
A candidate can have a high score and low confidence. That means the result may be promising, but the bank is not strong enough for hard decisioning.
Why scores can change after rebuilding
Rebuilding a signature changes the brand standard for future scorings. Scores can change when:
- New references are added.
- Labels change.
- References are excluded.
- Weights change.
- Calibration selects a different blend.
Historical results remain historical. Compare scorings across time with awareness of which signature produced them.