Cortex is organized around a small set of product nouns. Understanding these makes every page in the app easier to read.
Workspace
A workspace owns brands, teammates, billing, usage, and API access. Members can work inside the same brand library. Admins and owners can manage team access and billing-sensitive settings.
Brand
A brand is the container for one scoring standard. It holds brand context, references, signatures, candidate uploads, scoring history, calibration, and diagnostics.
Create separate brands when the creative standard is meaningfully different. For example, two unrelated advertisers should be separate brands. A single advertiser with multiple minor edit variants usually belongs in one brand.
Reference
A reference is a historical ad assigned to a brand. Each reference is labeled as a winner or loser.
Winners are examples of creative response patterns you want to reproduce. Losers are examples you want to avoid. Both sides matter. A brand with only winners does not teach Cortex what the brand's weaker response pattern looks like.
References also have review fields:
- Review status: pending, approved, or excluded.
- Weight: a value from 0 to 1 that controls how much the reference influences the next signature build.
- KPI metadata: optional JSON for notes such as spend, CPA, ROAS, hook rate, or platform context.
Signature
A signature is the brand's current winner-vs-loser response pattern. It is built from eligible references and used when scoring candidates.
The latest ready signature is the one used for new scorings. Rebuilding a signature does not rewrite old results; historical scorings still reflect the signature that was active when they were created.
Candidate
A candidate is a new ad you want to evaluate. Candidate uploads are verified, then scored against the brand's latest ready signature. Candidates do not become references automatically. If a candidate later becomes a proven winner or loser, add it to the reference library intentionally.
Scoring
A scoring is one run for one candidate against one brand signature. It can be reserved, running, succeeded, succeeded partially, or failed.
A successful scoring produces:
- Final score.
- Brand winner-likeness.
- General attention.
- Whole-video and element scores.
- Confidence.
- Recommendations.
- Playback-linked brain activity.
- Element and network diagnostics.
- A heatmap when available.
Calibration
Calibration uses leave-one-out validation to test how well the reference bank separates known winners from known losers. It can select the best video/element blend for that brand. If there is not enough data for full calibration, Cortex uses the default blend and marks confidence accordingly.
Diagnostics
Diagnostics help improve the reference bank. They flag references that behave unlike their assigned label, have low event coverage, or may deserve a lower weight before the next rebuild.