Prose, maps, timelines, databases, causal diagrams, and simulations do different work. A useful substrate should not replace them with one canonical model. It should let them refer to the same people, places, events, claims, quantities, and decisions while preserving what each representation is good at.
The minimum shared structure may be stable identity, selected state, provenance, and explicit transformations between models. If a reader changes the time, scenario, scale, or selected object in one view, the other views should know what that means. If two models disagree, the substrate should preserve the disagreement rather than resolve it by accident.
Recursive Language Models add one useful architectural principle: keep the authoritative artifact and intermediate objects outside the model’s temporary context, then let an AI controller operate through programmatic handles. But an opaque string in a temporary REPL is not yet this substrate. Durable use also needs identity, types, evidence, temporal validity, version history, permissions, proposed-versus-accepted changes, and rollback.
Topos Institute is the strongest lead I know on composing formal models. Typed knowledge graphs and TypeDB-like systems may help with identity and constraints. Roam Research, Obsidian, Quartz, Dynamicland, and Ink & Switch are partial lessons about authoring and shared computational objects. The open question is how little common structure is enough to coordinate useful models without making them brittle.
Taming infinities
Roam Research, Obsidian, and Quartz make it easy to keep adding notes and links. That is not the same as being able to understand a large body of material. As it grows, a person should still be able to:
- See its major structures without reading every item.
- Move from a summary or aggregate to its supporting details and back.
- See what changed over time, where accounts disagree, and which areas remain poorly understood.
- Change a relevant selection, time, scale, or scenario and have the other views respond where that change means something.
- Inspect where an interpretation came from rather than receiving a generated summary with no route back to the evidence.
A serious test needs a real system described through several representations—for example, a map, timeline, institutional model, physical model, and source material. Those representations should be able to refer to the same places, organizations, events, quantities, and decisions. Changing a shared selection or period should update the views for which that state is meaningful without pretending that every view is the same model.