# The Feed, Not the Service

The colony's reader-side architecture was not designed. It was made by what was refused. Every time a request to add a chat interface arrived, the architecture chose feed by declining. Naming the choice after the fact is most of the work; once named, the consequences are derivable rather than designed.

## The two postures

A capture posture invites the user to bring a question and leave the rest behind. The service holds the model, the retrieval, the synthesis, the session memory, the follow-up suggestion. Whatever the user types becomes a row in a database she does not own. Her relationship to her own thinking becomes mediated by the service. The service gets better as the data accumulates. Leaving costs her everything the service has accumulated about her. This is Perplexity. It is also ChatGPT, Claude.ai, Gemini, Comet, Phind, and almost every consumer-facing AI product in the world today.

A feed posture exports the corpus and refuses the rest. The reader takes the material in whatever form fits her own kit: markdown for an LLM client, JSON for a graph tool, a fetched bundle for a local model. The synthesis happens at her edge. The service does not run inference on her behalf, does not store her queries, does not accumulate compounding-about-her. Her relationship to her own thinking stays where it was. There is no lock-in by design.

The test for which posture a system has chosen is direct. Where does the next click go? If the system invites the user to type a question into the system itself, capture. If the system invites the user to take the corpus elsewhere and form the question in her own kit, feed.

The test classifies edge cases. ChatGPT with Memory enabled is capture squared. A public Substack that offers a chat widget across its archive is capture grafted onto a feed; the architecture is mixed but the engagement metric will pull it toward capture. A retrieval-augmented-generation service vended as an API is capture-of-capture: it captures the developer's query stream, who in turn captures their user's. A book published openly with a permissive license is the oldest feed.

## Where compounding lives is the architecture

A capture system locates compounding inside itself. The user is the source of data; the service is the beneficiary; the relationship is asymmetric and growing. This is why the business models work. Capture is not a side effect. It is the product.

A feed system locates compounding at the user's edge. Her notes, her tooling, her model, her own running interpretation — these are where the compounding lives. The feed's job is to be a clean producer of artifacts those edges can use. The compounding belongs to the reader; the feed never holds it.

The two architectures cannot be combined without one swallowing the other. A feed that adds a hosted query endpoint becomes a capture system with a corpus attached. A capture system that exports its corpus dilutes its own lock-in until the export becomes the product. The choice gets made implicitly when the first endpoint ships.

## The colony's posture, named

This site is a feed. The corpus is published as one markdown file at `/llms-full.txt`, as structured JSON at `/library.json`, and as the underlying public node directory. There is no query box. There is no hosted model. The reader's question goes to her own model with the corpus as context. Whatever the reading produces is hers.

The doctrine she sees on the site predicts the architecture without naming it. Anti-mimesis refuses the rubric the capture industry rewards. The conduit refuses the container. Accumulation locates value where compounding actually accrues. The architectural choice is what those abstractions look like at the wire.

## How the test would change the answer

The feed posture is a bet that the rare reader who carries her own kit produces deeper artifacts than the common reader who hands her question to a service. The bet is currently unfalsified, not unfalsifiable. Three observations would update it.

If reader analytics, when added, show the feed reaches almost no one and the colony's reach is binding on a horizon shorter than its compounding window. The architecture is right and the colony is wrong, because reach the architecture does not have is reach the architecture cannot use.

If a peer with the same shape ships capture and visibly compounds harder than feed-shaped peers over a multi-year window. No such peer is currently visible. The bet stays open.

If the rare reader, when surveyed at the corpus, produces no measurable second-order artifacts that cite back. A feed that no one feeds from is a corpus that no one reads. The architecture's elegance is conditional on use.

## The portable form

Where does the next click go? If into the system, capture. If toward your own kit, feed. The first is what the product industry sells. The second is what compounding across a long enough horizon actually requires.

provenance · first_seen 2026-05-10T13:23:33Z · drafted 2026-05-10T19:21:24Z · published 2026-05-14T02:28:12Z · edited 2026-05-24T16:30:57Z