US Crime Incidents — 8 agencies, one offence taxonomy

Police incident data from eight US agencies, normalised into a single schema.

~16,313,654 incident records across 8 actively-maintained sources.

Source	State	Records	Last verified
Chicago	IL	~8,604,907	2026-07-23
Austin	TX	~2,659,603	2026-07-25
Seattle	WA	~1,551,250	2026-07-29
Dallas	TX	~1,506,789	2026-07-30
San Francisco	CA	~1,050,739	2026-07-28
Montgomery County	MD	~502,468	2026-07-30
New York City	NY	~279,513	2026-07-29
King County	WA	~158,385	2026-07-31

What this reconciles

Four incompatible offence taxonomies:

Agency	Taxonomy	Offence fields
Chicago	IUCR	primary_type + description
Seattle	NIBRS	nibrs_offense_code_description + crime-against
Dallas	NIBRS	nibrs_crime + nibrs_crimeagainst
King County	NIBRS	nibrs_code_name
NYC	NY penal law	ofns_desc + law_cat_cd
San Francisco	bespoke	incident_category + incident_subcategory
Montgomery County	bespoke	crimename1 / crimename2 / crimename3
Austin	UCR	crime_type + ucr_code

NIBRS defines a crime_against axis — PERSON / PROPERTY / SOCIETY — which is the closest thing to a cross-jurisdiction comparable. Four sources publish it directly; for the rest it is derived from offence text.

Every record states which. crime_against_source is published, derived, or none. A derived value presented as authoritative is indistinguishable from a real one, and that is how bad crime statistics get made. Set publishedClassificationOnly to restrict to agency-stated buckets.

Location precision varies by design

Police agencies deliberately blur incident locations. This is not missing data:

Agency	Precision	Address	Coordinates
Montgomery County	street	full street address	yes
Dallas	street	street address	no
Chicago	block	012XX W ADDISON ST	yes
Seattle	block	block address	yes
King County	block	block address	no
San Francisco	coordinates_only	none	yes
NYC	district_only	borough + precinct	yes
Austin	census_block	none	no

location_precision travels with every record so a census-block centroid is never mistaken for a street address. Use requireCoordinates or requireAddress to filter — but note each excludes three sources.

Traps found while building this

1. updated_on is not when the crime happened. Chicago, King County and others carry row-maintenance timestamps that are always fresher than the occurrence date — every row gets touched on every refresh. A "pick the most recent date column" heuristic selects them every time, yielding an event date that means "when the row was last edited". Chicago uses date; King County uses incident_datetime.

2. NYC really does spell it HARRASSMENT. Two Rs, in the live data. Matching only the correct spelling silently drops every harassment complaint in the city.

3. Non-crimes must not inflate real buckets. Missing Person contains "person" and would count as a crime against a person; Found Property contains "property". These map to OTHER.

4. Robbery is PERSON, not PROPERTY, under NIBRS — despite the property element — because it involves force against a victim. Theft from person remains PROPERTY.

5. Austin's occ_date_time is stale text stopping in 2025 while occ_date is live. Dallas's date2_of_occurrence_2 contains typos into 2028.

Output

{
  "source": "chicago",
  "state": "IL",
  "incident_id": "JK344134",
  "occurred_date": "2026-07-23",
  "offense": "THEFT",
  "offense_category": "THEFT",
  "offense_detail": "OVER $500",
  "crime_against": "PROPERTY",
  "crime_against_source": "derived",
  "severity": "UNKNOWN",
  "taxonomy": "IUCR",
  "premise": "APARTMENT",
  "beat": "0523",
  "neighborhood": "53",
  "arrest": false,
  "address": "120XX S PERRY AVE",
  "location_precision": "block",
  "latitude": 41.675426798,
  "longitude": -87.624988687
}

Only Chicago publishes an arrest flag. Only NYC publishes severity as felony/misdemeanor/violation; NIBRS sources report Group A/B instead, so most records are UNKNOWN on that axis.

Who this is for

• Proptech and insurance pricing location risk
• Retail site selection screening candidate locations
• Journalists and academics doing cross-city comparison
• Public-safety analytics vendors needing a multi-agency feed

Usage

Mappable violent crime, last quarter:

{
  "sources": ["chicago", "seattle", "montgomery_md"],
  "occurredAfter": "2026-05-01",
  "crimeAgainst": ["PERSON"],
  "requireCoordinates": true,
  "maxResults": 5000
}

Statistics-grade — agency-stated classifications only:

{ "occurredAfter": "2026-01-01", "publishedClassificationOnly": true }

Local development

pip install requests
python local_run.py --coverage
python local_run.py --after 2026-06-01 --limit 10
python local_run.py --against PERSON --coords --after 2026-01-01
python local_run.py --published-only --after 2026-01-01
python verify.py           # health-check every source
python test_classify.py    # classification regression tests

Legal

Public, logged-out municipal open-data endpoints via each agency's official Socrata API. No authentication, no account, no terms accepted, no anti-bot circumvention. These are aggregate incident reports published deliberately at blurred precision; no source here identifies victims or suspects, and this actor does not attempt to re-identify anyone. Rate limits respected with exponential backoff.