sparrowmap.com — a licence-plate camera network that the neighbourhood owns instead of a police department renting.

"And ye shall be brought before governors and kings for my sake, for a testimony against them." — Matthew 10:18

Volunteers point a camera at a public road from their own property. The camera does all recognition locally and sends a detection event, never video. The results land on a public map anyone can open with no account.

Start Sparrow.bat
  map          http://localhost:8150/
  transparency http://localhost:8150/transparency
  camera app   https://localhost:8151/app          <- note https

The hub ships with a simulated town (sources/synthetic.py) so the whole system is testable against known ground truth with no hardware attached. It is off by default; python hub.py --sim turns it on.

Running a camera? The one-tap browser camera needs no install — open the site and press Add a camera. For the full local pipeline (accurate vehicle classification, plate reading, and the trained government-vehicle head), see docs/DESKTOP_NODE.md.

A network like this can be built two ways and only one is worth building. If every plate is public and searchable, anyone can look up an ex-partner's week with photographs. That is not accountability, that is a stalking service with a civic-sounding name.

So there are two tiers, enforced in code rather than in a policy document:

A publicly owned vehicle, doing public work, on a public road, is a public record. A private person driving to work is not.

1. The gate is conservative and corroborated. A vehicle reaches the public tier only above 0.85 confidence and only when something other than the plate text agrees. A single OCR slip must never be able to publish a stranger. classify.py.

2. A private plate is destroyed in the image, not just in the database. This is the hole that sinks most designs: every plate reader photographs the plate, so a private tier that keeps the photo keeps the plate. SparrowMap pixelates and then bars the plate region before the JPEG is written. snapshot.redact_plate, gated on tier != "public" in snapshot.py - a government plate stays legible in the snapshot on purpose, because that photograph is the public record.

3. Snapshots are crops of the vehicle, not the frame. The pedestrian, the house number, the kid in the yard are not in a tight crop of a car's rear end. Full-frame storage is off by default and requires face blurring when on.

4. The hash key rotates. Private plates are keyed hashes so a car still re-identifies across cameras. The key rotates every 30 days, so trails cannot be joined across the boundary. This caps how long anyone can follow a private citizen including whoever runs the hub. privacy.rotate_pepper.

5. No video crosses the network boundary. Recognition is on-device; what arrives is a few hundred bytes and one still. There is no stream to intercept, subpoena or leak.

core.py         paths + the entire privacy posture as one config block
privacy.py      plate hashing, pepper rotation, retention, redaction policy
classify.py     police / gov / fleet / civilian, with evidence and a hard gate
snapshot.py     vehicle crop, plate redaction, provenance stamp
nodes.py        ed25519 node identity, enrollment, camera cone geometry
db.py           sqlite schema + queries
hub.py          stdlib HTTP, the API, the janitor, TLS listener
sources/
  synthetic.py  a simulated town with ground truth, for building with no hardware
public/         map, about, transparency, phone contributor
data/           sparrow.db, snaps/, pepper.json  <- pepper.json is the crown jewel
certs/          self-signed TLS (browsers refuse the camera on plain http)

/api/policy exists so an outside auditor can diff a deployment's claims against its behaviour without being trusted with access. A promise nobody can check is not a promise.

• The signal weights in classify.py are starting values, not measurements. They need calibrating against locally labelled footage before the confidence numbers mean anything. See ROADMAP.
• Government vehicles whose only evidence is plate text stay private, on purpose. That is rule 2 doing its job, and it costs real coverage.
• Phone submissions are one person's eyes. They are allowed to publish, because a human looking at a marked patrol car beats any classifier, but they are marked unverified and attributed to the submitting node.
• The keyed hash is only as strong as the pepper. A plate has ~10^8 possible values, so anyone who steals data/pepper.json can reverse every live hash. Retention and rotation bound the damage; they do not eliminate it.

The code is AGPL-3.0 (see LICENSE). In plain terms:

• You can run it, read it, modify it, and share it, freely.
• If you run a modified version as a network service, you must publish your source, including your changes. That is the point: a fork used to watch the public must itself be open to the public. A surveillance vendor cannot take this, close it, and sell it back.
• Derivatives stay AGPL. Nobody can relicense it or take it proprietary.

The bundled vehicle detector (public/vendor/yolo11n.onnx, Ultralytics YOLO11) is itself AGPL, so the choice is partly the dependency's and partly ours, and it is the right one for this project.

The licence covers the CODE, not the NAME. SparrowMap, the sparrow mark, and sparrowmap.com are the identity of this project, not part of the AGPL grant. You are welcome, encouraged, to run your own instance from this code, but give it your own name and logo rather than presenting it as SparrowMap. (Same principle as Firefox: the source is open, the name is not.)

Running your own regional instance is the design. A network nobody owns is many small servers, not one big one.