Every failure in this series was a design choice. Here is the design that fixes it. Each part already exists.
Six essays of failure invite a fair question: fine, so what would work? The honest answer is unusually short, because nothing in it needs to be invented. Every failure this series measured came from a specific design choice, and every fix below has already been tested by somebody in this story, most of them by the government itself. A real count needs exactly four parts.
The one national count the government runs asks police agencies to participate voluntarily, and this series watched what that buys: six years to reach its own 80 percent publishing bar, one year over it, and then back below.[src] [live] A voluntary count is a count that can quietly un-happen, and the agencies most likely to opt out are exactly the ones nobody else is watching.
So the first line of the spec is the shortest: reporting is a duty, not a favor. Whether a legislature gets there with a requirement, a funding condition, or something else is politics, and politics is not this essay's lane. The design requirement is only that the count cannot depend on being volunteered.
Five projects counted the same country and got five totals, and essay four showed why: about 110 people a year appear or vanish on one definitional choice, shootings only or all killings.[src] That is not sloppiness. It is what happens when the word "killed" is left for each counter to define alone.
A law that says "count the killings" without writing down what counts will simply buy a sixth incomparable number. The spec says: one public definition, covering the deaths in custody, the restraints, and the vehicles, not just the gunshots, plus a published crosswalk so every older count can be translated onto it. The disagreement between the databases was never really about facts. It was about words, and words can be fixed in statute text.
Here is the strangest fact in the whole series. The federal government already tested the civilian trackers' method, media review first, agency confirmation second, and published the result. Its own assessment had found the old Arrest-Related Deaths program captured, at best, 49 percent of law-enforcement homicides,[src] so it suspended the program in 2014 and piloted the open-source approach instead. In ten months the pilot's media review identified 1,348 potential arrest-related deaths, roughly 135 a month, and the study estimated about 1,900 for the full year.[src]
The deepest undercount in this series was not a database anyone forgot to fill in. It was the death certificates, the government's most official record of all, which misattributed 55.5 percent of police-violence deaths over four decades without anyone noticing from the inside.[src] Systems that fail loudly get fixed. Counting systems fail silently, so a real count has to be audited from outside, on a schedule, by people who do not run it.
The auditors already exist too. The three civilian lists agree with each other on more than 98 percent of the actual people,[src] which makes them a ready-made external benchmark: any year the official total drifts from the civilian consensus, someone has to explain why in public. And the certificate fix is narrower than it sounds, better training and coding practice for the medical certifiers, and a routine link between law-enforcement records and the vital statistics. A coding fix, not a new bureaucracy.
None of this is a position on policing. It is a filing-cabinet specification, the same kind of design reading this series has applied to every count it examined. Whether the United States wants to count the people its police kill is a values question, and a country answers that one for itself. How to count them stopped being an open question years ago: mandate, definition, method, audit. Four parts, all tested, one of them validated by the government's own pilot a decade back. The engineering is done. The count is a choice.