#Work_IN
It is August 13, 2026, 12:20. I had a serious problem with the Librarian module. Experiments confirmed the worst: the module could not build high-quality connections, and it was too early to talk about the hidden connections that the system absolutely needs.
I had previously rejected marker-based connections such as Work_In, Live_In, and similar labels. They would turn the system into a large, poorly controlled if/else machine built around markers. I tried to compromise. We do not always understand why we associate Monday with the color blue. How could I create something I do not understand myself? I could not.
I spent a lot of time thinking about the problem. I tried hidden connections, temporary marker connections, and eventually marker connections again, but none of it worked. The system stayed around 30–55% accuracy, which is not acceptable for the next mechanisms.
Then I had an exciting idea. Previously I trained LB like a normal neural network: it received an input and had to build the correct graph. But what if I removed that output? It worked! I completely removed the gold graph as a target concept and started training LB to build graphs independently, requiring only high-quality answers.
The connections between concepts are now formed by the neural system itself. That is the only, and not especially important, price for making Librarian much stronger than before. The improvement was immediate: metrics increased by roughly 20%, from 55% to 75% accuracy.
LB is now beginning to hallucinate because of this training approach, which violates one of the project's principles. I will keep watching it and work on removing the problem.
Thank you for reading! More notes are coming soon :)