Limitations
What this study cannot infer—and why those boundaries matter more than a polished chart.
Public-data selection
Atlas and OpenStreetMap coverage reflect what researchers, agencies, reporters, and volunteers can document. Missingness is not random. Camera counts are lower bounds, and vendor representation may be biased toward visible procurements.
Administrative inconsistency
Agencies may define detections, hits, duplicates, and reporting periods differently. The descriptive hit share is not comparable to sensitivity, specificity, or positive predictive value without verified ground truth and downstream outcomes.
No linked enforcement outcomes
The volume datasets do not connect each alert to visual verification, stop, force, arrest, recovery, charge, dismissal, or correction. Flockwatch cannot calculate the individual risk of an erroneous detention.
Incident ascertainment
Cases are more likely to become public when video exists, litigation is filed, advocates investigate, or a journalist is involved. The 13-row registry is not representative and cannot measure demographic disparity or vendor-specific risk.
Synthetic external validity
The OCR experiment uses one open model, one font, one fictional renderer, and an unusual small-token layout. Commercial systems may use different optics, multiple frames, proprietary models, jurisdiction priors, and human review. The values demonstrate a controlled information-loss pathway only.
Changing law and infrastructure
Statutes, contracts, sharing defaults, vendor features, and litigation change quickly. This publication is a July 2026 research snapshot, not legal advice.
No causal crime-effect estimate
The project does not estimate how adoption changes crime. A credible evaluation would require validated deployment dates, pre-specified outcomes, matched or randomized comparisons, spillover analysis, and controls for concurrent policy changes. UCR alone cannot supply that design.
Compact interpretation contract
- Atlas record ≠ camera.
- Mapped camera ≠ complete inventory.
- Detection ≠ unique person.
- Hit ≠ correct match or solved crime.
- Non-hit ≠ proof of innocence.
- Documented incident ≠ prevalence.
- Synthetic OCR result ≠ Flock accuracy.
These are not disclaimers added after the fact. They define the study's valid claim space.