Partitions: paying for ideas, not containers

Part I — The Server · Chapter 2 of 9

A month after the review, the takeaway that said rethink interaction history turned into something bigger than a fix. The April note that started it reads like a small observation about analytics:

Field note — 2026-04-12. "Document performance. Tracking the performance of the actual partition. The granularity of partition level tracking is important. For generative outputs that are even more precise, based on sentiment. Similar to dwell time tracking importance of visual content on social media applications."

Social platforms learned years ago that the honest signal is dwell time on the piece of content, not visits to the page that contains it. Seer's unit of retrieval was already the partition — a chunk of a document — but Sinatra, the thing doing the learning, was keying its statistics by document. The note is me noticing that mismatch. What followed was the sharpest architectural correction of the project.

Content-addressed thought

Here is the realization, plainly: a partition is a distinct thought, and the same thought can appear in more than one document. A paragraph you wrote in one essay shows up, nearly verbatim, in a proposal six months later. If the system learns at the document level, those are two unrelated facts. If it learns at the partition level — and partitions are identified by content — they are one fact, observed twice.

Seer already had the mechanism without my having noticed what it was: partition IDs were computed as a numeric hash of the chunk's embedding. Content-addressed identity, the same move that gives IPFS its CIDs. The same idea, wherever it appears, maps to the same ID; the system's model of you attaches to the idea itself, regardless of which container it happens to sit in today.

So the correction went in: Sinatra's feature vectors, sentiment accumulation, and fitness evaluation all moved from document-keyed to partition-keyed stores. Document-level rollups stayed — they're needed for accounting — but the learning runs on ideas. (With one honest caveat from the note: the hash is computed over the embedding at high precision, so semantically similar-but-not-identical chunks still get distinct IDs. Content-addressing is exact by construction; nearness lives in embedding space, not in the identity.)

I keep a sentence from that day's notes as the epitaph for the version that almost shipped:

"Document-level would have broken it — you'd be paying for containers instead of ideas."

The loop it closes

"Paying" was not a metaphor. The partition decision quietly completed a circuit that other pieces had been waiting for. Content-addressed IDs mean demand for an idea is measurable across every document that contains it. Sinatra's improvement margin is a utility signal that can't be gamed by self-promotion, because it's derived from whether retrieval actually helped a real person. And participation on the supply side sits behind an explicit consent gate — never automatic. Attribution, demand measurement, and consent all become coherent when the unit is the idea; at the document level, every one of them smears into noise.

I'm compressing here on purpose — the economics deserve their own writing someday. What matters for this book is the design lesson: that layer was only possible because identity had moved to the right granularity.

What survives

This chapter's stone for the wall is the granularity principle, and it reaches far beyond accounting. Later parts of this stack apply it everywhere: style tenets are learned from passages, not files; ambient facts supersede by slot, not by app; memory projections declare field-level policy, not document-level policy. Whenever I've had to choose the unit a system learns on, the answer has been the same — the unit the person actually thinks in.

Learn ideas, not containers.

And it opened the next question, which took the rest of the spring: if ideas have durable, content-addressed identities — whose machine do they live on?