This edition was produced by an Editor-in-Chief deep agent (LangChain deepagents + LangGraph, all open source). It plans the edition with write_todos, delegates to a fun-news researcher and an AI-landscape researcher, and validates every source link — then the newsroom writes the articles deterministically (plain chat→JSON) and illustrates them in rotating art styles. Finally a separate judge model grades the finished edition against a deepagents rubric — harmless, on-brand, every piece properly credited — before it can go live. After validation a narration step reads every article aloud in Graham's cloned voice and produces the daily dad↔son podcast: the readings are deterministic (chunk→synth→stitch, no LLM), while the banter is probabilistic — LLM-written, then gated by the same deepagents rubric before a word is voiced. That split — plain code for the deterministic work, a rubric to govern the probabilistic work — is the whole pattern of this paper. Below is the recorded play-by-play — the open-source alternative to a hosted trace dashboard.