FLOCKWATCH / RESEARCH DESIGN

Methods

A source hierarchy, explicit units, deterministic transforms, public-data provenance, and a clear account of what the study cannot infer.

public sources onlyreproducible Pythonno UCR proxy

Study design

Flockwatch combines four evidence streams:

  1. Deployment inventories — EFF Atlas of Surveillance and Eyes Off Indiana/OpenStreetMap-derived camera mapping.
  2. Administrative activity records — EFF California detection/hit files and public agency transparency portals.
  3. Documents — peer-reviewed studies, government evaluations, court opinions, statutes, audits, procurement records, public logs, and evidence-graded reporting.
  4. Local experiment — a deterministic string-collision model and 600 inferences on fictional plate crops using an open OCR model.

Source hierarchy

Claims prefer, in order: court opinions and statutes; official audits and government datasets; peer-reviewed research; agency-released records; university public-records research; first-person accounts with underlying documents; investigative journalism; advocacy analysis linked to records; and vendor material.

The hierarchy does not make government sources infallible or advocacy sources unusable. It preserves the source's institutional position so the paper does not mistake a customer survey for a causal experiment or a marketing case study for an independent effect size.

Units and denominators

TermMeaning in this studyNot interchangeable with
Atlas recordSourced agency/jurisdiction technology recordcamera
Mapped cameraPublicly documented camera objectcomplete installed inventory
DetectionPlate-read event as reported by agencyunique driver or vehicle
HitRecorded match to a hot listcorrect match, stop, arrest, recovery, conviction
SearchDatabase querydownstream enforcement action
IncidentPublicly documented case in purposive registryprevalence or risk rate

Why FBI UCR is excluded

The FBI's Uniform Crime Reporting program records offenses known to police and related arrest/clearance statistics. It does not record ALPR scans, retention, query purposes, stale hot-list entries, false alerts, verification failures, cross-border searches, or stops caused by bad matches.

UCR could be one outcome source in a separate causal crime study with validated deployment timing, credible comparison jurisdictions, spillover analysis, and pre-specified outcomes. Using it here as a direct measure of surveillance harm would be a category error.

Reproducible transformations

  • Public inputs are declared in sources/data_sources.json.
  • scripts/fetch_data.py downloads them and records SHA-256 hashes.
  • scripts/run_analysis.py creates all descriptive tables and labeled SVG figures.
  • scripts/run_ocr_lab.py renders the fictional images and produces every inference row.
  • scripts/build_paper.py builds HTML/PDF editions and synchronizes public downloads.
  • Eleven unit tests plus manuscript contract tests pin parsing, counts, collision behavior, deterministic rendering, and research disclaimers.

Ethics and privacy

The project does not query live ALPR systems, collect private movement histories, publish real plate-owner records, or attempt to identify drivers. Synthetic identifiers are fictional. Public audit-log findings are used at the aggregate or already-published incident level.

Download the provenance manifest → · Read the limitations →