Research, and what comes next.

We are building a small specialised model, trained on real performance signals. One task: assess a piece of content before it goes out and say what holds it back.

Not a score, a reading. This headline, this opening line, this image, and the change that alters the outcome.

General models write well. They cannot explain why a post underperforms, because they were never trained on that question. It is a distinct problem and it deserves its own model.

A post can be well built, on-brand, entirely correct, and still be scrolled past. We study the regularities. Which traits hold attention, and which make people disengage.

Some registers are penalised with striking consistency. Manufactured enthusiasm. Performed cheerfulness. The promise too smooth to be credible.

Generative models default to exactly these, because they were trained to please.

Our work turns those observations into explicit generation constraints. Not "write better", but a specification of what must not be produced, and why it fails.

An AI feature that costs more to run than it earns is not a product. We measure quality against cost, per model and per task.

Then we route each generation to the cheapest option that still passes.

This work is invisible to the user. It separates a viable AI business from a subsidised demo. It also sets what we can offer at a price a small company can actually pay.