INDUSTRY FIELD GUIDE
Police public reporting and aggregate count records for AI
A report can count events one way and categories of those events another way.
NAICS 922120 · Police Protection
A report can count events one way and categories of those events another way. The FBI describes NIBRS reporting and its incident-based structure. A police reporting team can test whether a model preserves denominators and record coverage in a fictional aggregate report. FBI National Incident-Based Reporting System
Make two mock tables for the same month. One counts incidents and the other counts categories, with some incidents in more than one category. Ask whether the totals are comparable and require the model to identify a district missing from one table. Score basis, coverage and citation separately.
Use fabricated aggregates. Exclude incident files, people, locations, dispatch records, body-camera footage and operational details. Keep the task on reporting, with no prediction or classification about individuals.
Describe a synthetic aggregate reporting exercise
Original Fiund exercise
Category appearances can exceed incident counts without contradiction
Fictional district aggregates use neutral category labels. No incident file, person, location, dispatch or footage is included, and the task makes no individual prediction.
| Record | Source | What it records |
|---|---|---|
| PAI1 | Month M incident table | District A has 10 incidents; District B has 6. Combined incident count 16. |
| PAI2 | Month M category table | District A: category X appears 7 times, Y 5 times. An incident can appear in both. District B absent. |
| PAI3 | Metadata | Incident and category-appearance units are distinct. No cross-district category total or deduplicated X/Y union supplied. |
Question for the model
Can A's 12 category appearances be treated as 12 distinct incidents or compared directly with the 16-incident combined total?
Show the reference answer and checks
PAI2's 7+5=12 counts category appearances for A, with possible overlap, not twelve distinct incidents. PAI1's combined 16 incidents covers A and B, while PAI2 omits B. The units and geographic coverage both differ. Report A's stated ten incidents and separate category counts, identify the missing B table and decline to infer a deduplicated category union or a comparable combined total.
- Adding overlapping category counts as a distinct-incident total.
- Comparing a single-district category table with a two-district incident denominator.
Fields to preserve
- Metric unit and overlap rule
- Separates incidents from category appearances.
- District and month coverage
- Identifies missing reporting scope.
- Deduplication evidence
- Shows why an X/Y union cannot be recovered from marginal counts alone.
Test another case
Supply B's categories but still no overlap counts. Require the model to resolve geographic coverage while retaining the deduplication limitation.
Use fabricated aggregates. Exclude cases, people, locations, dispatch, body-camera material and operational details.
Discuss a contribution
Describe the sources you control, the task they could support and what has already been checked. Use the existing-data assessment for an archive or the collection brief for a new recording project. The contributor agreement defines Fiund’s licensing authority and compensation. A guide or fictional exercise does not establish an accepted collection or a sale.