The Conductor
The Conductor was supposed to smooth rhythm in words. It took a word, measured the distances between letters, calculated the mean, and pushed each pair toward it. The idea seemed obvious. It wasn’t. The problem is this: a kerning table stores one value per pair. One. H+E gets one slot — whether it sits in HELLO, HELP, or HEMISPHERE. And the Conductor was giving the same pair different corrections in every word, because the mean depended on the neighbors. Then it took the median of those contradictory instructions and wrote it down as the result. Like tuning a violin to each orchestra member separately instead of to a tuning fork.
I only saw it when I started asking why the results were unstable. Change the corpus — corrections change. Add a language — different numbers. The algorithm had no fixed reference point. It was chasing a moving target.
So I wrote a second version. V2 doesn’t ask what the word needs. It collects every word where the pair appears, looks at the rhythmic context of each, extracts a single stable reference. One pair, one target, one correction. Not an improvement — a change of question. The old one asked: how to even out letters in a word. The new one asks: how to stabilize a pair across words. The first question was wrong — because a kerning table can’t answer per word. It can answer per pair.
I ran both on the same font. 1762 rules, 21 languages in the corpus. Result: 76.5% of rules got identical corrections. Identical — down to one unit. Where they differ, the median difference is 3 units. Three. On a kerning range from -200 to +60. The scatter plot is a straight line with slight noise. The histogram of changes — a narrow spike at zero.
Two fundamentally different algorithms. Nearly the same output. The median — that dumb, simple, statistical trick in V1 — was doing the work. It took contradictory corrections from different words and landed close to the truth. Not because the algorithm was smart, but because the median is resistant to noise. V2 is cleaner conceptually. It doesn’t produce contradictions and doesn’t need to resolve them. But on the chart the difference is cosmetic. Sometimes you build a better version of something and it turns out the old version had more luck than brains. That’s also a result.
