How we verify a municipality
People and AI agents act on zoning answers. We publish our measurements and their limits alongside our review process. A citation or an unchanged source file alone does not establish that an answer reflects all adopted law.
Measured accuracy — incomplete baseline
Measured 2026-10-19. Overall accuracy is not established. We randomly selected 385 of 6,450 published rules without replacement, using random allocation within city, topic and district-category strata. Small strata can have no selected rules. Of those, 101 could be scored against the supplied source documents; 284 remain unassessed. Missing selections were not replaced.
- Source agreement among assessed rules
- 84.2% (85/101)
- 95% Wilson interval, assessed subset only
- 75.8%–90.0%
- Confirmed errors among assessed sampled rules
- 16 of 101 assessed (385 sampled)
- Cohen’s κ for rule-answer type
- 0.696 (97 resolved pairs)
Error breakdown: transcription: 2; omission: 11; applicability: 2; stale: 0; citation: 1. Corrections do not replace the original results in this baseline.
The blind pass used an isolated extraction prompt with no access to the published values or citations. The historical drafting model is not known. A separate comparison checked the frozen rules against those primary documents, and flagged disagreements received GPT-family adjudication. Supported assessments did not receive a further independent adjudication. The sample size targets ±5 percentage points at 95% confidence only if the sample is fully assessed. Source availability is not random: this subset’s interval does not describe overall accuracy. Archived and codified sources can miss later adopted law. κ measures agreement on answer type, not correctness of dimensions or legal outcomes.
Unassessed: 234 source unavailable; 40 insufficient source evidence; 10 superseded-source applicability unresolved; 0 awaiting adjudication. Treating every missing selection first as wrong and then as right gives frozen-sample bounds of 22.1%–95.8%. These are missing-data bounds, not a population confidence interval.
Confirmed errors include incorrect or omitted dimensional qualifications, use applicability and citation errors. Prepared content corrections remain subject to the publication gates and human approval. Automatic resolution in Franklin and Rochester Hills is disabled after confirmed wrong-parcel matches. Gallatin’s previous-code pack is withdrawn while the newly adopted code is reviewed. Legislative and map validation remain incomplete across the coverage list.
Re-measurement is scheduled quarterly; next due 2027-01-19. Download measurement and method · Sample and supporting evidence
The review process
- Draft from the official source. Record the ordinance and read the relevant tables, definitions and exceptions. Mark unreadable or unavailable provisions as gaps.
- Independent substantive review. A separate pass checks district rosters, dimensional standards, applicability, use permissions, citations and claimed absences against primary sources. A full review has its own date; an automated source check does not advance it.
- Map validation. Test address resolution and compare GIS classifications with adopted map instruments. Disable automatic resolution when a mismatch is confirmed. The current program has not completed this legal comparison for every automatic city.
- Publication and regression gates. Require citations, complete staged imports and applicable review gates. Frozen facts and error-specific tests check releases. Passing those checks is not a measured accuracy percentage.
- Amendment monitoring. Check supported source documents and agenda feeds. Monitoring coverage is incomplete. An unchanged codified copy does not rule out a new ordinance; adopted amendments require substantive review.
What the dates and citations mean
Municipality answers identify the pack version, source-check date and substantive-review date when known. Unknown dates remain unknown. Source links let you inspect the text, but parcel-specific approvals, overlays, variances and later adopted law may change its application. Consult the municipality for a binding determination.