Start Here

1. What this study measured

Flockwatch is a public-data study of fault amplification in networked automated license-plate-reader (ALPR) systems—how uncertain camera reads, hot lists, national sharing, and human verification failures can combine into coercive police action.

Four evidence streams, all public:

Full design and source hierarchy: Methods.

2. What this study did not measure

Preserve these boundaries; they define the valid claim space:

• No private Flock access and no live plate queries
• No UCR proxy for ALPR harm or effectiveness
• No national false-positive rate and no Flock proprietary accuracy estimate
• No causal crime-reduction claim from the public files used here
• OpenStreetMap / Atlas figures are lower bounds or adoption inventories, not complete installed inventories or market share

Details: Limitations.

3. The fault chain (not a single bad camera)

The motivating public case and the lab treat harm as a chain, not a solitary sensor glitch:

1. Source record loses information (incomplete hot-list entry)
2. Camera / OCR loses different information
3. Matching treats lossy strings as equivalent
4. Network distributes the alert
5. Interface compresses uncertainty into a hit
6. Officers treat the alert as self-authenticating suspicion

The laboratory reproduces a mechanism—100 synthetic identifiers collapsing to one key after middle-token loss, plus open-model OCR under controlled degradation. That is not a vendor accuracy score. See Error lab.

4. The three-denominator problem

Numbers on this site answer different questions. Do not add them or collapse them.

Interpretation rule and figures: Measured findings.

5. Where the 29 documented cases fit

The incident registry holds 29 evidence-graded public cases (grades A–B). They illustrate failure classes—OCR/sensor error, stale records, sharing failures, insider misuse, profiling queries, security exposure—not how common those failures are.

• Documented incident ≠ prevalence
• Ascertainment is biased toward video, litigation, audits, and journalism
• Pending charges remain allegations until adjudicated

Use cases to understand pathways; use inventories and activity files for scale.

6. What governance breakpoints mean

Perfect sensors are impossible. High-stakes systems need checks that break the chain before uncertain data becomes force: full-identifier matching, independent officer verification, short retention, restricted sharing, warrants for historical search, purpose limits, auditable access, and independent evaluation.

The study maps twelve breakpoints across capture, matching, access, and remedy. A control that only shortens retention does not stop a dangerous real-time stop; a verification rule that officers never apply does not contain automation bias.

Framework: Twelve breakpoints.

7. How to read the rest of the site

Suggested paths by role—each link stays inside existing routes:

Compact contract (from the validity page): Atlas record ≠ camera · mapped camera ≠ complete inventory · detection ≠ unique person · hit ≠ correct match · non-hit ≠ innocence · documented incident ≠ prevalence · synthetic OCR ≠ Flock accuracy.

Read limitations in full → · Reproduce the analysis → · Full paper →