The problem was never a shortage of geographic data — it was the inability to trust any single number on a map. This platform makes provenance, currency, and use-rights part of the database itself, so trusted answers are produced by construction rather than by discipline. The scaffold is built, the trust controls are proven on a live database, and all core architecture decisions are resolved.
The Situation
A Governed Platform
Not a database — a system that knows where every dataset came from, how current it is, who owns it, and whether AI may use it. One spine instead of scattered spreadsheets.
Trust Is Structural
A file that loaded without error is not a number you can defend. Here, bad data is rejected at the database and unapproved data cannot reach an AI — enforced, not hoped.
Proven Scaffold
Schema applies clean on Postgres 16 / PostGIS 3.4, guards reject bad data, history resolves correctly, and the AI gate holds. Ready for the first real load.
Proven, Not Asserted
- Bad data is refused at the door. Integer-coerced FIPS codes, a duplicate "current" assignment, and orphan rows are all rejected by the database itself.
- History resolves to the right boundary. A 2023 event joins to the chapter that owned the county then — not today — fixing a whole class of silent map errors by construction.
- The AI can only see what is approved. A dedicated read role reaches the serving layer and is denied every base table — 5 of 5 isolation tests pass. The approval flag is now an enforced gate, not a label.
Decisions — All Resolved
| ID | Decision | Resolution | Status |
|---|---|---|---|
| D-005 | Canonical cloud database | Supabase Postgres + PostGIS; local analytics in DuckDB. | RESOLVED |
| D-006 | Enforce the AI serving wall | Read-only role, serving layer only — built & proven. | RESOLVED |
| D-007 | Canonical store for geometry | PostGIS canonical; GeoParquet archive-only. | RESOLVED |
The Path, In One Line
Build the governed scaffold → prove the guards and the AI gate (done) → load the master geography into the spine → load the first topic dataset and log the run → generate the first end-to-end county report → hand to adversarial review → only then add more sources.
- Next step: register the master-geography source and load county boundaries + chapter assignments into the spine.
The Ask
Approve standing up the canonical Supabase database so the first real geography load can begin. Everything upstream is built and proven.