Writing conventions for kb/notes/ (theoretical register)

Register

This collection operates in the theoretical register. Documents here make transferable claims about what is true — mechanisms, principles, and general arguments that should hold across systems, not just in this one.

The quality goal is reach: the most general formulation the argument supports, with boundaries mapped explicitly. A note with reach compresses many situations into one explanation, making bounded context work harder.

Title conventions

Claim titles by default. Name the note like a claim, not a topic: "structure enables navigation without reading everything" rather than "thoughts on structure." The title should be something that could be true or false.

Composability test: the title should work as prose when linked — since [title](./title.md) or because [title](./title.md) should read naturally.

Claim strength test: the claim should be contestable. "Continuous learning is substrate-independent" fails — nobody would push back. "Continuous learning can happen outside of weights" names the thing people actually doubt.

Exceptions. Don't force a claim title when it feels strained. Common exceptions: - Multi-claim specs and frameworks (no single claim subsumes the content) - Definitional notes (the title is the term being defined) - Indexes (navigation hubs) - Seedlings not ready to assert a clear claim

If using a claim title, add the title-as-claim trait so review gates can check the promise.

Note composability

A note should be linkable from other notes without dragging irrelevant context. The title-level prose test checks whether the link text can work in an argument; the note-level composability test asks whether the body is focused enough that another note can use it as a premise without inheriting unrelated claims, examples, or local context.

If a note has one strong claim plus a second cluster that future links would not want to import, split the cluster into another note or move it to kb/work/ until its role is clear.

Reach as quality discipline

Aim for notes that explain why, not just record what works. Quick tests: - If you changed one premise, could you predict what changes in the conclusion? - Would the insight still apply in a different domain? - Could someone say exactly how the explanation is wrong, not just that it's incomplete?

Notes that only record "X works" are adaptive — useful but brittle. Notes that explain why X works have reach. Reach is a goal to move toward, not a gate every note must clear.

Formulation constraint. The theory's claim — title and opening argument — must be statable in general terms, even when derived from a specific system. If you can't state it without referencing a particular system, it's not yet a theory.

Outbound linking conventions

When linking FROM this collection:

To register Appropriate relationships Notes
Theoretical (same register) since / because / contradicts / extends / qualifies The claim-traversal graph. These are the argumentative links that make traversal-as-reasoning work.
Descriptive (reference, agent-memory-systems) evidence / derived-from / exemplifies Theories cite descriptions as observations and evidence. The theory must stand without any single description — if it can't, it's still a description, not yet a theory.
Prescriptive (instructions) evidence (rare) Theories rarely need to cite procedures. When they do, it's usually to note that a practice exists, not to depend on it.

What does NOT belong here

  • Descriptions of how a specific system works → descriptive register (kb/reference/ or kb/agent-memory-systems/)
  • Procedures, conventions, how-to guidance → prescriptive register (kb/instructions/)
  • Raw captures without frontmatter → text type, any collection
  • Work in progress → kb/work/ (workshops)