Hermiona works. I’m not continuing it.

I ran the full pair matrix on two Liberation fonts. 4056 inferences. Five minutes on CPU. Zero extraction errors. MAE 14.3 on the sans, 18.1 on the serif. The model sees that serif needs stronger kerning — on its own, no hints. It recognizes the classic hard pairs. T over lowercase, A diagonals before V. It can loosen — 20% of predictions are positive values. It sees that g+j in serif needs space because the descenders collide.

The interesting part: Liberation Sans has 55 pairs in its kerning tables. Hermiona proposes sensible values for 1829. Thirty-three times the coverage. And it’s not garbage — the patterns check out. Stems over lowercase, diagonals, extenders. It works. And that’s exactly why I’m not continuing it.

Hermiona learned from the work of 716 designers. People who sat for hours over letter pairs and decided: tighter here, looser there, leave this one. Every value in the training set is someone’s decision. Someone’s sensitivity. Someone’s years of experience.

A font being open source doesn’t mean you can take its kerning, pack it into a dataset, and train a model that replaces that work. The license says “you can use it”. It doesn’t say “you can teach a machine to do what I do”.

I look at what companies like Anthropic do. They take people’s writing. They take people’s code. They take people’s art. They pack it into a model and sell it as a service. They write “developed in collaboration with Claude” as if it were a partnership. It’s not a partnership. It’s extracting value from other people’s work without asking.

I was doing the same thing. Smaller scale, but the same mechanism.

I need to find another way. Maybe synthetic shapes generated parametrically — mimicking letters, but with no designer’s history behind them. Pure geometric kerning is out — I tested seven algorithms on hundreds of fonts. The best one barely pulls a −0.75 correlation. Not enough. But something between raw geometry and stealing other people’s decisions must exist. I don’t know what yet. But I know it’s not this.

Hermiona works. The code stays. The documentation is done. But training on other people’s work — that’s over.