The universe, built bottom up
The second edition is written from a universe of concepts, claims and evidence, and the universe is built bottom up: one local graph per carried source document, each document keeping its own vocabulary, with bridges between them added later as authored decisions. Nothing here is book text. This is the material the book's own graph will connect to.
Layer 0 · the frozen bytes: the 21 carried sources under
/v1/docs/sources/, each hashed in the freeze manifest, so an anchor into one is stable forever.Layer 1 · per-document local graphs, this section: the document's own dictionary, claims with their support state, hypotheses, objectives, demonstrations and asserted relations, every one anchored to a verbatim quote. The build refuses to ship an anchor whose quote is not found in the frozen bytes, so extraction cannot cite words that are not there. Every layer 1 node means only: this document says this, here.
Layer 2 · the bridge layer, not started: cross-document edges, authored with a note, recording where documents name the same idea differently and where they disagree. Divergence is preserved as a finding, never merged away.
Layer 3 · the book's universe, not started: the six node families of the dev pack's ADR-3, connecting down into layers 1 and 2 as evidence.
The sources, and where extraction stands
| Document | State | Yield |
|---|---|---|
| Thinking in Graphs: Meaning Through Connectivity dated 2026-02-05 · the cornerstone; the book's subtitle is this document's subtitle | extracted · pilot | 22 concepts · 27 claims · 8 edges · 8 known uses · the folder · PDF, 14pp |
| the other 20 carried sources | queued | the pilot settles the method first; each will get the same treatment |
The document of one node
The experiment of brief 24, programmatic phase: pick a node and read the document that grows from it — everything the anchored extraction verifiably holds about that one concept, composed in document structure with no authored prose. The picker ranks every node by how many links it carries, which makes the richness of each concept measurable before any writing happens.
The rules this section holds itself to
- One extraction file per document, and the dictionary, taxonomy, thesaurus and ontology views are projections of it. Four authored artefacts per document would be four things that can disagree; one file with four views cannot.
- Every node is anchored to a verbatim quote in a named section, and the build verifies the quote against the frozen bytes on every release. A hallucinated quote fails the build; a misread of a real quote is what review is for.
- Coverage is total by construction: every section with prose either yields anchored items or is recorded as deliberately empty with a reason. A recorded empty section is a finding; a silent one is a hole.
- Truth is deferred. Layer 1 records what a document says, including its own evidence state (demonstrated, argued, declared). Whether it is right is layer 3's question.
- The extractor is an author too. The extraction is the agent's reading, marked as such, and it ships with the anchors a reviewer needs to check it. Undefined-but-used terms are recorded as named absences.
For an agent
The machine surface is data/universe.json: every node, edge, alias and distinction, with resolved byte offsets into the frozen sources and the SHA-256 each anchor was verified against. Each document's standalone folder under docs/ holds its source copy, its extraction and its cross-references (crossrefs.json, rated against usage-model.json). Treat a layer 1 node as a statement about a document, never as a statement about the world. If you are reviewing an extraction, use its PDF and read nothing else; come back item by item against the node ids. If you are building or changing anything interactive, first read the immediate-connection register: it names the experience these viewers are built towards, maps which patterns already exist, and carries the checklist a change should pass.