How we built the model — and how we present what it can't do.

Cortexa is trained on within-channel pair data: every training example is a pair of creatives that ran on the same channel with overlapping audience definitions. We deliberately avoid cross-channel comparisons in the training set because they confound model fit with platform-distribution effects.

Validation uses a strict channel-disjoint split: no creator appears in both train and val. The headline number you see across the marketing pages — 66.7% pair accuracy — is measured on 84 such channel-disjoint pairs.

Every creative is scored on four pillars — Hook, Hold, Algorithmic Fit, and Brand Lift — each on [0, 1]. The overall Winner Score is the harmonic mean across the four, not the arithmetic mean. The harmonic mean drags the composite down sharply when any single pillar fails, which prevents the model from rewarding creatives that win on a single axis (e.g., a strong hook that retains nobody).

Predictions are conditioned on a demographic cell — country, gender, age band, device, daypart — because we observe and report substantial within-creative variance across cells. Holding the cell fixed is what makes the Winner Score directly comparable across variants in a single comparison.

On Cortexa Create, every result includes a three-segment horizontal bar. The first segment is the share of variance we can attribute to brain-pattern predictability. The second is the share that, by our own measurements, is best explained as platform-algorithmic chance — distribution effects we cannot move. The third is unknown noise. We bound these segments empirically; we do not claim the unknown share is zero.

Each result is jittered slightly within an empirical confidence band so identical re-runs aren't bit-equal. This is deliberate: the model itself is not noise-free at the granularity of a single pair, and the UI should not present it as if it were.

After you run a comparison, Cortexa will ask you 7 days later which variant actually won. Your answer becomes a labeled row in our training set, weighted by the verification level (manually attested vs API-verified). No model checkpoint ships to production until it beats the prior version on the held-out actuals slice — Brier score on confidence, plus per-metric mean absolute error.

Cortexa is not a substitute for live testing. We predict within-channel pair winners, not absolute campaign outcomes. If you only ever ran two ads, our prediction is the right tool. If you're choosing between dozens of variants across cells with very different distribution dynamics, you should still A/B test the top candidates.