About the labels

1. 1

   Luna reads the brief

   A language model reads the text of the brief and writes a breakdown: the idea, the pains, the core message, the hook, the script structure. It only restates what the text says; it does not judge.

2. 2

   Jev labels the breakdown

   A second model reads the breakdown and gives every axis of the genome a probability: story shape, hook device, desire, mechanism and so on. The label of an axis is its most probable value, and that probability is the confidence.

3. 3

   The verifier checks the reading

   A third pass checks how much of the breakdown the brief's own text supports. A breakdown that does not pass is kept, but the brief is marked as needing review. The exact model versions used for a brief are listed in its Analysis section.

4. 4

   Embeddings group the ideas

   Each brief also becomes a vector of its meaning. Briefs with close vectors form idea families, and the nearest earlier brief of a direction decides whether a new idea repeats.

The genome is the fixed list of axes every brief is labelled on. Labels change only when the genome version changes, and every analysed brief says which version it was read with.

“Uncertain” is a label too: no value reached the confidence bar, so the brief is counted but not placed on that axis.

A root idea counts as a repeat when its creative premise (the move and the concrete subject, not the product) is at least — similar, as a cosine on the idea vector, to an earlier unrelated root idea of the same direction. Two briefs in different tasks are text near-duplicates from —, as a cosine on the vector of the brief text: the same words written twice. Both cut-offs belong to the embedding model on the right.

Each cut-off was measured on this corpus: it is the similarity that only a small share of unrelated pairs reaches (0.70 % for an idea repeat, 0.33 % for a text near-duplicate). People have not yet labelled pairs of briefs to check it against.

Use the repeat rate to compare directions and months with each other, not as a count of real repeats: the cut-off is applied to the closest of hundreds of earlier ideas, so part of what it flags is chance.

The team's own dictionaries (core message, pain point, visual format, headline formula) stay as Airtable holds them, with the model's suggestion beside them. A value a person sets always beats a model's.