Building a Machine That Reads Luhmann's Handwriting Good Enough
Six AI models, 73,715 cards, one day. 15.6% error rate. What happens when you stop aiming for perfection and start aiming for useful.
Articles on mapping the most famous knowledge system in history.
Six AI models, 73,715 cards, one day. 15.6% error rate. What happens when you stop aiming for perfection and start aiming for useful.
From 2 images/second to 23/second. How a disk-full error led to parallel downloads, a machine migration, and a deep dive into German copyright law.
When AI can't read 45-year-old German handwriting, the surrounding cards can help. A multi-agent system that lets the network read itself. CER from 163% to 27%.
A 10-layer synaptic model, myelination scores, and an 8-dimensional diamond layer. Why this architecture works on the most famous knowledge system in history.
After mapping 73,715 cards and 18,289 cross-references, eight lessons emerge that apply to anyone building a personal knowledge system.
We mapped 58% of Luhmann's Zettelkasten. Here's what becomes possible when we have the rest.
The first network analysis of all 73,715 cards reveals surprises about his thinking that 70 years of scholarship missed.
The hidden API, the download script (80 lines), and why the archive was always open — nobody took it.