FIELD NOTE · SUPPLEMENTAL SET
A million alerts and 161 false flags: California's plate-reader year
ISSUEDJULY 11, 2026DRAWN BY THE NIGHTLY SWEEP
Oakland's plate readers logged 638,747,333 plates last year and threw 1,099,837 hotlist alerts. The department left the biggest alert class switched off, because it had nobody to send. In Los Angeles, the LAPD's Inspector General counted 161 alerts on cars that were not stolen. Same technology, same state, same year, and both columns are true.
Start with Oakland, because the figures are the department's own annual report: 293 cameras produced 638 million reads and 1.1 million alerts, more than 620,000 of them for stolen vehicles or stolen plates. The alerts are 0.17 percent of the scans. The stolen-plate share is 0.09 percent. The department said in that report that it lacks the staffing to respond to those alerts.
The biggest alert class on that network is the one nobody is answering.
The other column landed in July. On July 11, 2026 the LAPD said it would not renew its Flock Safety contract, after its Inspector General published a 98-page audit of the plate-reader program. That audit covered all three systems LAPD runs, pole-mounted cameras from Motorola and Flock plus cruiser cameras from Axon. Across two months, August 1 to September 30, 2025, it found 161 alerts where the plate matched and the car turned out not to be stolen, and over the same two months the department recovered 337 stolen cars.
Reporters added those two numbers and divided. The one-in-three error rate in the coverage is that arithmetic, and 161 alerts and 337 recovered cars are not the same unit. The report's own totals for the window are bigger and stranger: more than 210.5 million reads, 5,911 plates tracked, no action taken on 4,575 of them.
An integrator selling into this market is selling both columns, whether or not you name the second in the bid, and the buyer, a city council, a shopping-center owner, an HOA board voting on a camera at the gate, has only ever been shown the first.
The read is not the weak link.
A plate can be read perfectly and the alert still be wrong, because an alert is only as good as the list behind it, and the list is a separate system with its own currency problem. This is the ordinary shape of a signal chain. The part everybody benchmarks is rarely the part that fails, and the handoff between two systems nobody owns jointly is where the failure lives.
The LAPD audit is that failure written out. Most of those 161 cars really had been reported stolen at some point, then recovered, or the report on them cleared, and the plate stayed on the hotlist anyway. LAPD's own response blames delays by another jurisdiction or by the owner in clearing a plate off the list.
So the number that would settle this is the one almost nobody publishes: how often an alert is right, counted after somebody checked. LAPD's Inspector General published a piece of that: one department, two months, 161 alerts on cars that were not stolen.
Which is where this stops being an ordinary false alarm. A nuisance trip on a detector empties a building and irritates everybody in it. A false stolen-vehicle alert puts an officer behind a car whose driver doesn't know why.
Accuracy isn't what LAPD said ended it, though. The department cited who owns the data, what happens to it once it is collected, privacy, security, and sharing. None of that is a camera problem. The audit produced the false-flag count, but the terms sheet is what the department named on the way out, and those are two conversations with two different people inside the same customer.
Los Angeles was running close to 2,000 plate readers anyway, and Oakland had 293.
So the honest demand story isn't a wave. It's enforcement activity and owner fear, both real without a wave behind them. California's CHP-led Organized Retail Crime Task Force has run 4,577 investigations since 2019, made 5,134 arrests, and recovered 1,620,803 stolen items worth $74.6 million, counted through April 2026. In 2024 alone: 879 investigations, 1,707 arrests, 676,227 items, about $13.5 million recovered.
Both halves of that, the crime decline and the enforcement totals, come out of one May 13, 2026 release from the Governor's office. A press release summarizes agency data and is not the data itself, and the 2024 breakdown came from trade and local reporting. Worth knowing before either number ends up in a proposal.
A falling statewide property-crime rate and a store that got cleaned out are not the same fact, and the buyer only has one of them.
California's cannabis rules don't invite a licensee to consider video surveillance. Title 4, Section 15044 specifies it: continuous recording, 24 hours a day, at a minimum of 15 frames per second, at a minimum resolution of 1280 by 720, with 90 calendar days of retention. Coverage is enumerated too, entrances and exits, limited-access areas, point of sale, and every area where cannabis is handled, with enough clarity at a retail register to make out facial features.
A spec that carries its own measurement conditions is rarer than it ought to be.
The licensee is a buyer on a deadline instead of a buyer with an opinion. The camera package is a condition of opening, and it comes due at the end of a buildout, after everything upstream has run long. Ninety calendar days at 15 frames per second also means the storage gets sized on day one or discovered later, when the head end fills and the oldest footage rolls off on its own.
The filing that carries it is a tenant improvement or a change of use, and it hits the public record weeks before anyone calls it a security project, in Irvine or Santa Ana or anywhere with a permit counter. PermitFeed scores that record nightly for security and low-voltage integrators, because the buildout is in the file and the camera order isn't.
Section 1798.90.51 puts two duties on all of them: reasonable operational, administrative, technical, and physical security safeguards, and a publicly available usage and privacy policy.
The policy has to name things. Authorized purposes, the personnel and job titles who can reach the data and the training they have had, how access gets monitored, what restrictions apply to selling or sharing or transferring it. Most camera scopes carry less specificity than that. It also has to name the responsible official, the measures that exist to correct errors, how long the data is kept, and the process for destroying it once the retention clock runs out.
The error-correction clause is the one that maps onto the audit, and it has been in the statute since 2016.
Buyers stopped waiting for a regulator to enforce it. The San Francisco Standard counted 82 Flock contracts terminated across 28 states between August 2021 and May 2026, and 39 of those came in the first five months of 2026 alone. That is a press tally, not a registry, and the best public count anyone has. Where terminations have been reported in detail, the trigger was a data-sharing violation, unauthorized federal access, or a governance failure.
82 FLOCK CONTRACTS TERMINATED ACROSS 28 STATES, AUGUST 2021 TO MAY 2026 39 OF THOSE TERMINATIONS FELL IN THE FIRST FIVE MONTHS OF 2026
That reads like a market shrinking. From inside a bid it looks more like a market picking up a specification it never had, written by the accounts that walked.
What holds up is dull and most of it is contractual. Retention set to a number and enforced, an access list with job titles, a written answer to who owns the data and who eats a breach, and a documented path for a wrong record to get corrected before it fires an alert. All of that is already in Section 1798.90.51, in language a city attorney can be walked through line by line.
The audit ran two months, one department. In it, 161 alerts told an officer the car in front of him was stolen when it wasn't. None of the people in those cars picked the vendor, wrote the retention policy, or had a vote on the pole the camera is mounted to.
Start with Oakland, because the figures are the department's own annual report: 293 cameras produced 638 million reads and 1.1 million alerts, more than 620,000 of them for stolen vehicles or stolen plates. The alerts are 0.17 percent of the scans. The stolen-plate share is 0.09 percent. The department said in that report that it lacks the staffing to respond to those alerts.
The biggest alert class on that network is the one nobody is answering.
The other column landed in July. On July 11, 2026 the LAPD said it would not renew its Flock Safety contract, after its Inspector General published a 98-page audit of the plate-reader program. That audit covered all three systems LAPD runs, pole-mounted cameras from Motorola and Flock plus cruiser cameras from Axon. Across two months, August 1 to September 30, 2025, it found 161 alerts where the plate matched and the car turned out not to be stolen, and over the same two months the department recovered 337 stolen cars.
Reporters added those two numbers and divided. The one-in-three error rate in the coverage is that arithmetic, and 161 alerts and 337 recovered cars are not the same unit. The report's own totals for the window are bigger and stranger: more than 210.5 million reads, 5,911 plates tracked, no action taken on 4,575 of them.
An integrator selling into this market is selling both columns, whether or not you name the second in the bid, and the buyer, a city council, a shopping-center owner, an HOA board voting on a camera at the gate, has only ever been shown the first.
Both columns come off the same camera
The chain is four links and you can hold it in your head: a camera photographs a plate, software reads the characters, the read gets matched against a hotlist, and an alert goes to whoever subscribes to it.The read is not the weak link.
A plate can be read perfectly and the alert still be wrong, because an alert is only as good as the list behind it, and the list is a separate system with its own currency problem. This is the ordinary shape of a signal chain. The part everybody benchmarks is rarely the part that fails, and the handoff between two systems nobody owns jointly is where the failure lives.
The LAPD audit is that failure written out. Most of those 161 cars really had been reported stolen at some point, then recovered, or the report on them cleared, and the plate stayed on the hotlist anyway. LAPD's own response blames delays by another jurisdiction or by the owner in clearing a plate off the list.
So the number that would settle this is the one almost nobody publishes: how often an alert is right, counted after somebody checked. LAPD's Inspector General published a piece of that: one department, two months, 161 alerts on cars that were not stolen.
Which is where this stops being an ordinary false alarm. A nuisance trip on a detector empties a building and irritates everybody in it. A false stolen-vehicle alert puts an officer behind a car whose driver doesn't know why.
Accuracy isn't what LAPD said ended it, though. The department cited who owns the data, what happens to it once it is collected, privacy, security, and sharing. None of that is a camera problem. The audit produced the false-flag count, but the terms sheet is what the department named on the way out, and those are two conversations with two different people inside the same customer.
Nobody bought a camera because crime went up
The demand story the trade tells is a crime wave. The preliminary FBI numbers for 2025 don't support it: across agencies covering 83 percent of California's population, violent crime fell 9.94 percent against 2024, property crime fell 14.35 percent, and murder fell 16.51 percent, with burglary and larceny-theft still declining.Los Angeles was running close to 2,000 plate readers anyway, and Oakland had 293.
So the honest demand story isn't a wave. It's enforcement activity and owner fear, both real without a wave behind them. California's CHP-led Organized Retail Crime Task Force has run 4,577 investigations since 2019, made 5,134 arrests, and recovered 1,620,803 stolen items worth $74.6 million, counted through April 2026. In 2024 alone: 879 investigations, 1,707 arrests, 676,227 items, about $13.5 million recovered.
Both halves of that, the crime decline and the enforcement totals, come out of one May 13, 2026 release from the Governor's office. A press release summarizes agency data and is not the data itself, and the 2024 breakdown came from trade and local reporting. Worth knowing before either number ends up in a proposal.
A falling statewide property-crime rate and a store that got cleaned out are not the same fact, and the buyer only has one of them.
The one occupancy that can't open without it
Most security work is a preference with a budget attached. One category isn't.California's cannabis rules don't invite a licensee to consider video surveillance. Title 4, Section 15044 specifies it: continuous recording, 24 hours a day, at a minimum of 15 frames per second, at a minimum resolution of 1280 by 720, with 90 calendar days of retention. Coverage is enumerated too, entrances and exits, limited-access areas, point of sale, and every area where cannabis is handled, with enough clarity at a retail register to make out facial features.
A spec that carries its own measurement conditions is rarer than it ought to be.
The licensee is a buyer on a deadline instead of a buyer with an opinion. The camera package is a condition of opening, and it comes due at the end of a buildout, after everything upstream has run long. Ninety calendar days at 15 frames per second also means the storage gets sized on day one or discovered later, when the head end fills and the oldest footage rolls off on its own.
The filing that carries it is a tenant improvement or a change of use, and it hits the public record weeks before anyone calls it a security project, in Irvine or Santa Ana or anywhere with a permit counter. PermitFeed scores that record nightly for security and low-voltage integrators, because the buildout is in the file and the camera order isn't.
The privacy policy is part of the spec now
California settled the question of who counts as an operator, and it isn't only the police: Civil Code Section 1798.90.5 et seq., added by SB 34 in 2015 and effective January 1, 2016, reaches any ALPR operator, public or private. The HOA is an operator. So is the business improvement district, the self-storage yard, and the retailer with a camera on the driveway.Section 1798.90.51 puts two duties on all of them: reasonable operational, administrative, technical, and physical security safeguards, and a publicly available usage and privacy policy.
The policy has to name things. Authorized purposes, the personnel and job titles who can reach the data and the training they have had, how access gets monitored, what restrictions apply to selling or sharing or transferring it. Most camera scopes carry less specificity than that. It also has to name the responsible official, the measures that exist to correct errors, how long the data is kept, and the process for destroying it once the retention clock runs out.
The error-correction clause is the one that maps onto the audit, and it has been in the statute since 2016.
Buyers stopped waiting for a regulator to enforce it. The San Francisco Standard counted 82 Flock contracts terminated across 28 states between August 2021 and May 2026, and 39 of those came in the first five months of 2026 alone. That is a press tally, not a registry, and the best public count anyone has. Where terminations have been reported in detail, the trigger was a data-sharing violation, unauthorized federal access, or a governance failure.
82 FLOCK CONTRACTS TERMINATED ACROSS 28 STATES, AUGUST 2021 TO MAY 2026 39 OF THOSE TERMINATIONS FELL IN THE FIRST FIVE MONTHS OF 2026
That reads like a market shrinking. From inside a bid it looks more like a market picking up a specification it never had, written by the accounts that walked.
What holds up is dull and most of it is contractual. Retention set to a number and enforced, an access list with job titles, a written answer to who owns the data and who eats a breach, and a documented path for a wrong record to get corrected before it fires an alert. All of that is already in Section 1798.90.51, in language a city attorney can be walked through line by line.
The audit ran two months, one department. In it, 161 alerts told an officer the car in front of him was stolen when it wasn't. None of the people in those cars picked the vendor, wrote the retention policy, or had a vote on the pole the camera is mounted to.