Twenty-Three of Eighty-Three

Philadelphia's Citizens Police Oversight Commission asked the police department for a month of body-camera footage from the 22nd District. So little came back that the auditors added a second month, just to have enough to look at. Their count, published on July 8th: of eighty-three encounters that department policy required officers to record, twenty-three were recorded.

“Just 28% of all incidents that required recording were captured by the officers in this sample, which is very concerning,” the commission's senior analyst, Michael Tooley Jr., said.

He is right, and the concern is the obvious one. But there is a second thing in that sentence, and it is the one this post is about. Somebody was able to compute that number. Doing so required two records that disagree with each other, and almost nobody has both.

The audit is a join

A body camera is a second witness to an event the paperwork already claims happened. The incident report says an officer responded to a disturbance on a particular block at a particular hour. The video system says whether a file exists with that officer's ID on it, starting around that time. Two systems, two authorities, one event.

The commission's method was ordinary database work. Pull 125 incidents at random from April and June of 2025. Read PPD Directive 4.21, which specifies when an officer must activate, and sort the sample into incidents that required a recording and incidents that did not. Eighty-three required one. Then ask the video system how many of the eighty-three produced a file. Twenty-three.

Neither number is the finding. The finding lives in the space between them, and that space is only visible to someone holding both sides.

One department, one directive, four answers

The 22nd was not the only district Philadelphia audited. It was the fourth, and the worst by a wide margin.

Same city. Same directive. Same vendor, same hardware, same year. The rate at which an encounter leaves behind a durable record varies by a factor of three depending on which side of Philadelphia the encounter happened on.

These are small samples of individual districts across single months, not department-wide rates, and the commission does not present them as such. The spread is the useful part regardless. Any measure built on “did a recording exist” inherits a coverage rate that changes across a city boundary nobody drew for that purpose.

One methodological detail is worth sitting with. The commission expanded the 22nd District review to a second month because so few recordings came back from the first. The sampling frame had to bend around the missingness it was built to measure.

What the camera writes down

Public argument about body cameras is almost entirely about the video. Whether to release it, when, to whom, with whose face blurred. That argument has consumed a decade.

Underneath the video sits something much smaller and much more useful to anyone building software: a row. Device serial, officer identifier, start timestamp, duration, the category the officer assigned to the recording, whether the device was activated from standby so the pre-event buffer was captured, and on many systems a set of coordinates.

The commission measured two of those fields directly. Of the recordings that did exist, 88% were activated properly from standby, and 97% carried the correct incident category. Those are good numbers, and they are good numbers about a minority of the encounters.

A log of those rows across a department and a year is a structured, machine-readable census of police-public contact. It contains no faces, no audio, no addresses of victims. It is the cheapest transparency artifact any agency with cameras already owns. We are not aware of a US department that publishes it.

The shape of the problem will look familiar to anyone who read our piece on drone-as-first-responder, where the dispatch system and the records system tell different stories about the same call. Cameras add a third system that agrees with neither.

The request wall

Suppose you ask for the footage anyway. New York City Comptroller Brad Lander audited how that goes at the NYPD, releasing the results on October 31, 2025. His office reviewed 5,427 requests for body-camera footage under the state Freedom of Information Law.

Of those, 4,591 blew past the 25-business-day standard. That is 85%. The average response took 133 business days. More than a thousand took longer than 200.

Between 2020 and 2024, requesters filed 355 appeals when the department refused or ignored them. The city granted 344.

Ninety-seven percent of appeals succeeded. The average wait was more than half a working year.

The same audit looked at the footage itself. Across 7,797 videos, officers activated late or stopped early in 1,436 of them, about 18%. And of the videos that should have accompanied 911 dispatches, 4,319 were missing outright. Roughly a third.

“The whole purpose of body-worn cameras is transparency,” Lander said. “Failing to provide the footage as required by law fundamentally undermines the whole purpose of program.”

The exemption structure here is the same one we walked through in what crime data APIs suppress. Video carries far more privately identifying content than a text incident record, so it draws far more redaction work, and redaction work is what produces a 133-day average.

The retention clock

California Penal Code section 832.18 sets the floor for nonevidentiary body-camera data at sixty days. After that an agency may erase it. Across states the routine-footage minimum runs from about two weeks to about six months. Footage tied to an arrest, a use of force, or an open complaint gets flagged and held much longer, often a year or two.

Retention is set by storage cost, not by evidentiary logic. For a hundred-officer department, published estimates put annual program cost somewhere between $110,000 and $290,000, and the single largest lever on that number is how long video is kept. Every extra month multiplies the bill.

The metadata row costs almost nothing to keep. It is text. Whether agencies keep it past the video it describes is a question we have not seen answered in any published policy, which is itself an answer of a kind.

Not only an American problem

The Victorian Auditor-General's Office in Australia examined Victoria Police's cameras and found that in March 2021 officers activated in 83.6% of the instances where they were required to. Better than any Philadelphia district except the 24th.

The more interesting finding sits next to it. VAGO reported that Victoria Police had no consistent way to track officers' camera use and did not know its own overall compliance. The 83.6% figure exists because the auditor built it. The force could not have told you.

That is the pattern in every case here. The compliance rate is not a statistic an agency maintains and occasionally shares. It is a statistic that comes into being when an outside body does a join, and it stops existing when they stop.

Building against a record you cannot read

Body-camera data will not appear in a crime feed. It still changes what the feed means, and there are concrete things to do about it.

1. Never read missing footage as a missing event. In the 22nd District sample, sixty encounters happened and generated reports. The absence is in the camera system, not the world.
2. Request the log, not the video. A metadata export carries almost none of the identifying content that drives redaction delay, which means it should clear a records office far faster. Ask for fields, date range, and format explicitly.
3. Treat a camera rollout as a structural break. Deployment dates change officer behavior and change what gets written down. If your time series crosses one, mark it, the same way we mark the LAPD handoff gap.
4. Do not rank agencies on compliance figures from separate audits. Philadelphia sampled incidents against a written directive. VAGO ran device telemetry for a month. Those two methods do not produce comparable rates, and a leaderboard built from them measures the auditors.
5. Write the retention floor next to your request latency. If the second number is larger than the first, you are requesting something that may not exist by the time anyone reads the request.
6. When a join fails, store the failure. An incident number with no matching recording is a row worth keeping. Enough of them, over enough months, is the only version of this measurement anyone outside an oversight commission is going to get.

Nothing here bears on whether crime is rising or falling. Violent crime is down 5.9% and murder down 17.3% in the first half of 2026 across the 590 agencies reporting to the Real-Time Crime Index, and no camera policy explains any of it. The point is narrower. A record exists, it is generated automatically, it describes contact between police and the public, and nobody publishes it.

Five thousand communities have Flock cameras reading every plate that passes. We wrote about that asymmetry in one-way surveillance last spring. Here it is again, wearing a different device, pointed the other way, and switched off 72% of the time in one district of one city.

The cameras in the 22nd District were on the officers the entire time. They were charged, assigned, and docked at the end of every shift. Sixty encounters produced no file, and the department's own directive says those files should exist.

The only reason anyone knows they do not is that a commission pulled a hundred and twenty-five reports and checked them one at a time.