Appendix: the field notes

Appendix

The dated blockquotes threaded through this book come from real working notes, lightly trimmed and never rewritten. Here they are in order, with where each one landed.

Dated

• 2025-10-23 — Seer begins. The project's start date; no note survives from the day itself, only the repository. Anchors chapter 1.
• 2026-03-06 — the candid review. A cold architecture review of Seer: Sinatra's park→sentiment→train→infer loop and its market-indicator features; Oracle's protocol decoupling and self-correcting trust edges; the debts (cold start, a hardcoded relevance threshold, per-turn retrains, a managed backend as a single point of failure) — and the statement of the core claim itself. Woven through chapter 1; its takeaways drive chapters 2 and 3.
• 2026-04-12 — the partition insight. Dwell-time as the model for granularity; partitions as content-addressed thoughts that recur across documents; the attribution-and-consent loop that only closes at idea granularity. Woven through chapter 2.
• 2026-05-09 — the mesh. Two-level HNSW over gossiped tag embeddings; the rare-topic recall problem ("summarization destroys specificity"); the three compounding fixes and their cost on personal hardware. Woven through chapter 3.

Undated

• The moat note. "The commodity phase of AI infrastructure is ending…" — personalization depth as the next competitive dimension; "they remember what you said, they do not learn how you think"; the personalization is the file; "the humans aren't context for the model." Quoted in chapter 4 and in full argument in chapter 9.
• The equilibrium question. A systems-design sketch whose surviving contribution here is one question — how a system keeps competition alive while it drives growth. Carried forward, unanswered, at the close of chapter 9.
• The wearables note. The far horizon: camera-less wearable embeddings turning the network into "a living nervous system." Mentioned exactly once, as horizon rather than plan, in chapter 9.