The Page Layer Goes Away

For weeks the work had been the same move: make the machine enforce what it promises. Then the vault filled up with real material—manuals, books, a year of links—and the app had to survive being used. What did not survive was the page.

What the Pages Were

compile was the fourth step, and it turned clusters of claims into wiki pages: a title, prose, tags, a place in the sidebar. It was the reason a claim could be right while the page about it was stale, the reason lint carried exceptions, and the reason the reader had two kinds of knowledge to tell apart. On a vault of a few hundred sources it produced pages nobody opened.

A page is an argument. A cluster is an association. Rendering the second as the first means inventing a thesis the data never had—so the step was replaced by one that only groups.

What Replaced Them

Four steps now, named for what they do: capture → outline → distill → cluster. The first three are per source and strictly ordered; cluster is the vault-level step, runs only while no document step runs, and uses no model at all. It writes titleless memberships: no prose, no title, no tags, no page. A cluster is a cache, rebuildable from the claims, and it is visible in exactly one place—as the peers of a claim.

That deletion removed a vocabulary: no branch, no concept, no compiled page, no page-level dirty state. Three layers and one association remain: sources/ immutable, wiki/ human, atoms/ the claims, clusters/ the association. The words follow: claim, topic, source, outline, cluster, finding, owed.

The Wiki Is a Source

The wiki did not disappear—it changed owner. A wiki page is Markdown a human writes, with a title, prose and tags, and saving it writes a source representation that runs through capture, outline and distill like anything else. The human keeps the one artifact a machine should not author: the statement of what a thing is and why it matters.

Saving a page also fans out. Its links become sources of their own—a project description with a GitHub link produces a snapshot, an outline and claims without a second action. The wiki is not decoration on the corpus; it is an input to it.

The Numbers

The vault in daily use, at the time of writing: 277 sources (29 PDFs, the rest Markdown and URL snapshots), 150,000 claims, 9,851 clusters, 17 wiki pages, 30 saved Ask conversations, 20,265 canonical topics and 467 wiki tags.

The last two are the honest part. At roughly seven claims per topic, the namespace has outgrown reading: topics are free names, the model invents them per document, and curation merges what it can. A naming lock—asking once per document for the topic set—is the open design question, and growth makes it worse, never better.

What It Feels Like

No search box, no tree: seven lenses, Wiki, Claims, Sources, Attention, Bookmarks, Ask, Status. A claim opens as a document with its sources, its citing pages and its peers; Sources have addresses and show every file they own; Attention is the queue of what is wrong. Bookmarks are collected, not understood, and say so.

What Is Still Open

The list is short and real: the topic namespace, the granularity of a cluster, trash without a retention policy, and CLI verbs for the document steps so a vault can be fed headless. No MCP surface and no sync in v1. None of that is missing function; it is the next round of work on a tool that already runs a real vault every day. On the studio the project is available now, and that means what it should: a working state, not a finished one.

The page layer went away because nothing needed it. The pipeline was the product. The wiki survived because a human still has to say what something is—and the machine’s job is to keep that true.

The removed page layer is the chapter’s real subject: a step that rendered associations as arguments, and thereby invented a thesis the data never had. What remains is three layers and one association—immutable sources, human-written wiki pages, claims as the truth, and titleless clusters as a rebuildable cache. The final numbers are a diagnosis rather than a scoreboard: 150,000 claims against 20,000 free-named topics is a namespace problem, and the open question is whether curation can repair a vocabulary that grows per document.