Somewhere in England or Wales last year, a person died behind a locked door. Someone had to find them. Someone had to phone a family. Then, months later, that death became a row in a spreadsheet published by the Ministry of Justice, and in the version of the table we can read, the row does not say who they were, how they died, or how long they had been inside. It says 18.0, under a column heading called "Window" (Source: Ministry of Justice, Safety in Custody -- Deaths_in_prison_custody_1978_to_2024_accessible -- Table_1_10).

That is the whole entry. Eighteen point zero, for the year 2024, the most recent year available.

Now zoom out, because the framing here is the point. The Safety in Custody series is one of the most morally serious datasets the British state produces. It runs from 1978 to 2024, 47 years of recording the fact that some people sent into the care of the prison service do not come out alive. Every entry is a human being who was, at the moment of death, the direct responsibility of the government. That is not a policy abstraction. That is a duty of care with a name attached.

Which is exactly why a column called "Window" with no unit is such a problem. Is 18.0 a number of days? A percentage? An average? A smoothing window used to calculate a trend line? The decimal point tells you it is almost certainly not a count of people, because you cannot have eighteen tenths of a person. Beyond that, the accessible table as published does not tell the reader, and if you cannot tell what a number measures, you cannot hold anyone to account with it.

Here is the strongest honest reading against my own irritation. Statistical publications are built for statisticians first, and this figure is very likely defined clearly somewhere in the accompanying notes, technical annex or the main bulletin that sits alongside the table. The "accessible" version of a file is often a stripped-back, screen-reader-friendly rendering of a much richer workbook, and stripping tends to remove exactly the header rows and footnotes that carry the definitions. So this may be a formatting casualty rather than a decision. It also means my read of 18.0 as an average or a rate is inference, not fact.

But the effect on the public is the same either way. A person who wants to know whether prisons are getting safer opens the file, finds a number with no unit, and closes it again. Good news gets lost in that gap just as easily as bad news does. If deaths in custody fell last year, that is worth knowing plainly and loudly, because it means something changed for the better in a system where change is slow and hard-won.

My view, and I'll own it: for a dataset that counts deaths, plain-English labelling is not a nice-to-have. It is part of the duty of care. The families reading these tables are not analysts. They are people trying to find their relative in a number, and "Window: 18.0" gives them nothing to hold.

Publish the unit. It costs one line of text.