Two Notebooks Lost · The code was mostly right. The experiments were already lost.

I deleted two notebooks of audio research during this project, but not at once. Notebook 1 failed after four days on April 30. I wrote down what had gone wrong and started over that day. Notebook 2 failed for a different reason on May 6, less than a week later. Between them, the notebooks represented a little more than a week of nearly continuous agent work. The second had already produced two long blog posts that I put online. Three people read them. I had also asked friends to complete a listening study in their free time. The study was already in front of two people. One of them, my regular Phish-going friend Patrick, had spent an evening on it. I was planning to talk about the results at a music festival beginning two days later. All of it came down because each notebook contained a silent setup decision that invalidated what followed. The decisions were different. One crossed an information boundary. The other changed what the experiment was measuring. In both cases, most of the…