Every count in this dataset can be absent for two different reasons, and the difference is load bearing. A null means no located source establishes the value. A zero means a source established that the value is zero. Collapsing them turns "nobody reported this" into "this did not happen", which is the failure mode that makes public-record datasets misleading.
The gap is large
| Reported count | Unknown | Established zero |
|---|---|---|
| People apprehended | 36 | 7 |
| People charged | 61 | 5 |
| People convicted | 88 | 1 |
Sum those columns as though the nulls were zeros and you would report a conviction figure that is wrong by most of the cohort. The nulls are not small and they are not evenly spread: the later a stage of the legal process, the more often it is simply not published.
Where else it shows up
73 of 186 infrastructure items record no quantity. The action is established, the scale is not.
On successor records, the seized-again field is a three-state value: 19 established yes, 26 established no, and 16 carrying nothing at all. Those three are stored separately rather than merged into a boolean, precisely so the third does not get read as the second.
Dates work the same way. Where only a month or a year is known, the date field stays null and a separate precision field carries the partial value. No record gets a day invented to make a column look complete.
Two aggregates that must never be added
Reported counts come from official aggregates. Named counts are derived from the individual person records in this index. They answer different questions and they are never summed. A release reporting a hundred and forty two people charged, alongside seventeen people named individually, does not produce a hundred and fifty nine. It produces two figures that overlap by an unknown amount.
How the site renders an unknown
An unknown is rendered as words, never as a blank cell and never as a zero. If a page says the figure is not established, that is the record speaking rather than a formatting failure. The same rule governs the return question: a takedown with no successor record is never presented as having stayed down.
10 coverage gaps and 11 excluded candidates are published for the same reason, so the decisions can be checked rather than taken on trust. The full counting rules are on counting incidents, and 1361 individual corrections are listed in corrections and updates.