Status and freshness

What is loaded, what has actually been checked, and what is knowingly missing.

An index that hides its staleness is worse than one that admits it, so the gaps below are published in the same place as the numbers.

Resolver accuracy

The jurisdiction resolver is scored against 0 probe points taken from the county's own zoning layers — the source that already carries the right answer.

The first row is the one that matters: when the resolver reports high confidence it has never been wrong. Every error falls inside the low-confidence band, and every one of them was within 141 m of a jurisdiction line — annexation lag, since cities annex continuously and Census boundaries update yearly. Near a city limit, no boundary source is certain, and the resolver says so rather than guessing.

Coverage

Jurisdictions

19 — all resolve; 2 have case records, 1 have code text

Land-use cases

11,849 (11,739 county, 110 Duluth)

Cases with a source link

11,598 of 11,849

Cases with a parcel PIN

10,804

Code sections

858

Code tables

18, of which 1 verified

Meeting documents

0 (— pages)

Build data source

supabase (postgrest)

Snapshot taken

2026-08-13 20:02:27 UTC

Known gaps

• 17 of 18 code tables are unverified. Only Table 2-B has been checked against the rendered source page, and even its merged PUD and CBD rows remain unreliable. Dimensional standards are the most queried content in the corpus and the easiest to get silently wrong.
• 0 applicant name pairs await human review. They are queued rather than merged, so developer case counts are lower bounds.
• Duluth minutes are 57–69% scanned with no text layer. Where OCR is used the text is stored separately and marked as a reconstruction — it is never presented as a quotation of the record. The OCR pass has not been completed, so decisions and vote records are under-captured.
• Duluth case fields are agenda-mined, not database-sourced. The city runs no case tracker, so these rows come out of agenda and minutes PDFs by pattern matching, and the field-level quality is poor. Measured across all 110 of them:
  • 68 have a “location” containing no street number at all, 15 of which are plainly not addresses (as presented., {J})
  • 58 have a “request” that is just the word ORDINANCE
  • 24 have an applicant field that ran on into the applicant's mailing address, and 9 have none
  • only 16 carry a zoning district, and 74 an outcome
  Treat each of these as an index entry pointing at a PDF, not as a structured record. The linked document is the authority; the extracted fields are a finding aid.
• 16 municipalities have boundaries but no corpus. They resolve correctly — the resolver's job is complete for the whole county — but there is no code or case data behind them yet.
• Meeting body text is not published here (12.5 M characters). Only document metadata is indexed.

Legal posture

Markdown version of this page →