Methodology

This site is generated by an automated pipeline. There is no human reviewer. Articles are collected from public feeds, classified by a language model, clustered into incidents, and published on a nightly schedule; name redaction and evidence tiering are applied by fixed rules, not by editorial judgement on the day. Nothing here is an opinion piece; it is a sorted index of other people's records. What that costs you is set out under Known limitations, and every row links to the source so you can check it against the original.

Collection

News coverage mentioning Flock Safety, automated license plate readers, or ALPR is collected daily from public RSS feeds, time-boxed back to the company's founding year (2017). Each article's full text is extracted and classified into one of a small set of categories (crime solved, abuse or misconduct, false positive or misidentification, other). Duplicates are dropped by URL.

Audit

Every event is independently re-read by a language model that extracts the primary entity, the jurisdiction, the technical role the system played, and the best available outcome on record, then searches for corroborating sources. It assigns an evidence tier:

• Tier 1 — official or judicial. A court order, statutory audit, government release, or docketed filing exists.
• Tier 2 — multi-outlet or council. Several independent outlets, or a city council record.
• Tier 3 — unadjudicated or PR. A single outlet, a police press release, or a vendor statement, with no adjudication.
• Debunked or dropped. Later reporting contradicted the original claim.

Incidents

Events describing the same matter are clustered into incidents by entity and jurisdiction token overlap, so one arrest covered by five outlets counts once.

Known limitations

• Categories, tiers, and notes are machine-drafted. They are consistent, not infallible. Always read the source.
• Coverage skews toward what gets reported: successes are announced by police departments, harms surface through litigation and audits. Treat counts as a map of the record, not a measure of reality.
• Publication date is the article date, not the event date.
• An unfinished case usually stays unfinished here. A charge is recorded when it is reported; the outcome is only recorded if later coverage surfaces and the pipeline finds it. Any row whose disposition is still open after 90 days now says so on its face — how long it has been, and when the pipeline last searched for an outcome. Read “charged” as “reported charged on that date”, never as a current status, and follow the source.
• Names of people classified as accused are shown as initials. Initials reduce exposure; they do not make a person unidentifiable when the surrounding facts are specific.
• Roughly 0 deployment entities have no parsable state.

Corrections

If a row is wrong, the fix is a better source. See About for how to send one.