How to play: cards appear below as the camera reads cars. HOTLIST + high confidence = CONFIRM it. Low confidence lookalike (0/O, 8/B) = REJECT it. Ordinary driver = SHRED it fast β if the yellow retention bar empties, you kept an innocent person's data too long. Watch for Mission Creep cards.
The claim: "A first-party system would be prohibitively expensive!" Build a 9-camera city-owned system for Cabot. Pick one in each row. Beat the lease.
| City-owned build | Lease forever | |
|---|---|---|
| Year 1 | $21,861β$37,611 (standard $26,361) | $27,500 / yr $137,500 over 5 years |
| 5 years | $53,829β$69,579 |
Real numbers panel: these are Corey's hostile-QA'd planning estimates (Nov 17, 2026), not a vendor quote. Unpriced lines β connectivity, CJIS-compliant hosting, PoE power, and staff admin time β are disclosed unknowns; pricing them is the Citizens Commission's first job. Per-item prices in the game above are planning allowances (camera ~$2,000 owned from the reported Quincy, MA route; commercial plate software $500/camera/yr from the same route) β the totals table is the sourced comparison.
No trick code, and no downloads β the first version of this panel borrowed the Tesseract OCR engine from a CDN, and if that CDN is blocked the panel dies, which rather proves the wrong point. So this version carries its own OCR engine inside this one file: a template-matching reader written in plain JavaScript (the classic OCR technique β segment the serial band into characters, then match each character's pixels against rendered templates). It generates a brand-new plate drawn to look like a real Arkansas plate β sky-blue-to-white gradient, the Crater of Diamonds diamond in the center, maroon βArkansasβ on top, βThe Natural Stateβ at the bottom, current ABC 12D format (Arkansas skips O and Q β so does the generator) β with an invented serial, because publishing real plate data is exactly what this game is against. Then the built-in engine reads it. No server, no CDN, no upload. It works offline, from a file, forever.
Honest caveat: this plate is flat, clean, and straight-on β the easy case, even with the diamond behind the serial and the real Arkansas layout. Real roads add angle, night, dirt, and speed β which is an argument about cameras, placement, and confidence thresholds, not about paying rent forever on plate-reading software. Production open-source stacks (OpenALPR, FastALPR, CodeProject.AI) use plate-trained models and run locally for $0 in license fees.
The objection: "Fine, it's cheap. But can a city keep this data safer than a vendor can?" Fair question β and the answer is not "trust us." The answer is custody: who holds the data, who can reach it, how long it exists, and who finds out when something goes wrong. Play it both ways.
Same camera. Same read. Very different blast radius if anything goes wrong β because one copy lives in a network thousands of organizations can reach, and the other stops existing.
Security is choices, and choices have prices. You cannot afford everything. Neither can a real city β that argument is the honest part of this round.
Each incident fires at the vendor system and the city system at the same time. The vendor side has no buttons. That is not a game mechanic β that is the contract.
Waitingβ¦
Waitingβ¦
The figures: 5,000+ law-enforcement agencies on the vendor's network (the company's own reported number); 7,000+ agencies and organizations able to reach shared data (ACLU of Massachusetts, Oct 2025, from network audits); 12M+ searches by 3,900+ agencies in logs obtained by the EFF (Dec 2024βOct 2025). Cabot's own sharing settings live in the vendor's portal β the city has to ask the vendor who can see the city's data. The exposure math below is illustrative: reach Γ days a record exists. And the standing caveat: custody makes security possible, auditable, and enforceable. It does not make it automatic β as incident 1 will happily demonstrate if you skip the patching.
The camera says HIT. Papers, please. Plate readers match numbers β they misread characters, and they are notoriously sloppy about which state issued the plate. The alert in front of you is a lead, not a fact. Your whole job is the five-second check: state, every character, vehicle. About a quarter of these alerts carry the right number on the wrong state's plate.
The receipts: Aurora, Colorado, August 2020 β Brittney Gilliam and four children were held at gunpoint, face-down on the pavement, because a plate reader matched her Colorado SUV's number to a stolen Montana motorcycle. Wrong state, wrong vehicle, right number. Aurora settled for $1.9 million. The Institute for Justice has documented at least 27 erroneous plate-reader encounters since 2018 β and attributes most of them to human error, not camera error. Independent testing (IPVM) has found readers misidentify a plate's state roughly one time in ten. Every one of those failures was preventable by the check you just did: state, characters, vehicle β before the stop, not after.
"It's just a license plate. It doesn't track anyone." Drive one stranger's ordinary week β no crimes, no warrants, nothing to hide β and watch what the cameras write down. Then we flip the screen around.
Every search of the plate database leaves a log line. Somebody is supposed to read them. Today, that's you. Thirteen searches from one small department. Approve the legitimate ones. Flag the abuse. This is not a hypothetical problem β misuse of plate databases by officers is one of the most documented failure modes of these systems, and the only thing standing against it is a person reading this log.
In the vendor model, this log lives on the vendor's servers, and the city sees what the vendor's portal shows it. In a city-owned system with a Citizens Commission, this log is the city's own record β and reading it is a public job with a name attached. Oversight is not a feature you buy. It is a person, reading, on the record.
Cabot has 9 plate cameras. Here's a town. Click intersections to place your 9 β put them where you'd catch stolen cars: the highways in and out, the main streets. Then run one ordinary day and count what your network actually records. The stolen car is out there. So is everyone else.
A simulation with toy numbers β 30 trips on a toy map. Scale it honestly in your head: a real town sees thousands of trips a day, 365 days a year, and at industry-typical rates well under 1 in 1,000 reads is a hotlist hit. The ratio is the point, and the ratio doesn't care about the size of the map.
A car gets reported stolen and goes on the hotlist. Then it gets recovered β and somebody has to take it off the list. That somebody is you, and the paperwork never stops. Clear the RECOVERED cars before a patrol unit finds them first and makes a "stolen car" stop on the owner. Leave the STILL STOLEN cars alone β pulling a live one off the list is its own disaster. Sixty seconds. The queue is not sorry.
Stale data is a documented cause of wrongful plate-reader stops β recovered vehicles that were never cleared have put innocent owners on the pavement in the Institute for Justice's case files. The software didn't fail in those stories. The list did. And lists are maintained by people, on purpose, or not at all.
Simple errand: drive from HOME (bottom left) to GRANDMA'S (top right). One condition β you never consented to be tracked, so find the route where no camera reads your plate. Click a highlighted intersection to move. Every π· zone you enter logs your plate, permanently, in somebody's database. Fewest reads wins. Zero would be a miracle.
The cameras on this map are placed the way real networks grow: not everywhere at once, just at the chokepoints β the bridges, the highway ramps, the one road across town. You do not need a camera on every corner to log every trip. You need them where trips cannot avoid.
Last year, your plate was read and searched like everyone else's. You'd like to know by whom. Reasonable. You have a right to ask. This is the asking. Every door costs days; some cost money; all of them are real doors that real people have stood in front of.
Same town, same nine cameras, same stolen-car problem. Leaseville rents the whole system from a vendor, the way most cities do it. Ownburg buys its cameras, keeps its data, and stands up a Citizens Commission. You make Ownburg's calls, one a year. Leaseville's are made for it β that's rather the point. Safety comes out about the same in both towns, because the cameras are the same cameras. Everything else is governance.
Money figures ride on the same record as Budget Boss: Cabot's $27,500/yr invoiced lease vs the QA'd city-owned estimate ($26,361 standard Year 1, ~$7,992/yr run rate, $53,829β$69,579 over five years, unpriced lines disclosed there). Trust and safety scores are illustrative β the shape is the argument, not the decimals.