Notes from DaedalusTechnical notes from the one-person AI studio of Siegfried Martens, on the systems it ships and how they were measured.https://s-martens.com/notes/2026-08-26T00:00:00ZSiegfried MartensForty-three thousand rooms and no serverhttps://s-martens.com/notes/forty-three-thousand-rooms/2026-08-26T00:00:00Z2026-08-26T00:00:00ZOrpheus serves a 43,567-room music knowledge graph with no backend, no database and no runtime code. Here is what that buys, what it costs, and the four places the approach actually hurts.Four sources that disagreedhttps://s-martens.com/notes/four-sources-that-disagreed/2026-08-26T00:00:00Z2026-08-26T00:00:00ZWikipedia, Wikidata, MusicBrainz and Spotify each have an opinion about who an artist is and what they play. They do not agree. Here is what I did instead of merging them, and the three ways identity broke anyway.Someone else's rulerhttps://s-martens.com/notes/someone-elses-ruler/2026-08-26T00:00:00Z2026-08-26T00:00:00ZI graded my own retrieval system on 115 questions I wrote myself. Then I ran it against three public benchmarks written by strangers, with predictions filed before the runs. Two agreed. The third found a real bug.The control did the workhttps://s-martens.com/notes/the-control-did-the-work/2026-08-26T00:00:00Z2026-08-26T00:00:00ZI ran three published benchmarks against a retrieval system in one day. Every headline number they produced was misleading until a cheap control corrected it. Four times out of four.The ruler came firsthttps://s-martens.com/notes/the-ruler-came-first/2026-08-14T00:00:00Z2026-08-26T00:00:00ZHow Pythia, an oracle over the Orpheus music knowledge graph, was built evaluation-first: recall@10 0.40 on day two, 0.89 at launch, and six times the measurement itself was wrong.