Building the Local Internet in Public

Where should the authoritative version of a business fact live?

The same business fact often appears in a dozen places with a dozen different values, and none of them state where the value came from.

Provenance is the missing field. Without it every copy of a fact looks equally authoritative to a machine.

A record that carries provenance and a timestamp on each field is more useful to machines than a page that carries none.

Model a record with per-field source and freshness, then render the same record in three different contexts.

Early modelling only. Field-level provenance makes the record noticeably larger and raises real questions about who may edit what.

Whether owners will maintain field-level detail over time.

What is the smallest set of fields worth verifying first?

Does a relationship carry information that a page does not?

A listing says what a business is. It rarely says how it relates to the city, the service category, the events around it or the people behind it.

Relationships appear to encode the part of local context that pages leave implicit.

Explicit relationships improve context for both human browsing and machine retrieval.

Build a small city-scale graph and compare answer quality against flat listings.

In progress. No documented comparison yet.

Which relationship types actually change retrieval outcomes.

Which relationship types matter most for local retrieval?

What would an agent actually ask a local information layer?

Agent queries tend to be scoped by place, category and constraint rather than by keyword.

That suggests a tool surface, not a search box, is the right shape.

A tool surface built around entity, geography and constraint fits agents better than a search box.

Draft a tool surface and write example agent transcripts against it.

Concept stage. No functioning public MCP service exists yet.

The right unit of response for an agent.

What is the right unit of response: a record, a list or a graph fragment?

Why do AI systems cite some local sources and not others?

Structure, authority and freshness keep appearing together in the sources that do get cited.

If those three travel together, citability may be designable rather than purely reputational.

Citability is a property of the information, not only of the publisher.

Publish identical facts in structured and unstructured form and observe retrieval.

Not yet run.

How stable citation behaviour is across models and over time.

Can citability be described well enough to be designed for?

Should geography be a filter or a first-class part of the architecture?

Most local systems treat geography as a filter applied after the fact rather than as structure.

Filtering discards the hierarchy that would have given the answer its context.

Nation → state → county → city → entity as native structure gives cleaner context for machines.

Model the hierarchy explicitly and route entity retrieval through it.

Early modelling. The hierarchy is clean; the edge cases at boundaries are not.

How to represent places that cross administrative boundaries.

How should places that cross administrative boundaries be represented?