How this registry works

One entry = one finding, one source, one exact location. Nothing else.

We do not investigate. We do not verify claims. We say who established a finding, in which document, and exactly where to look. The reader checks without believing us — that is the whole design.

What we never do

We never rate a claim no "true", no "false", no verdict. The page's invisible markup carries no rating field, because there is nothing to put in it
We never strip a figure of its caveat if the source says "estimated", "at least", "reported by", so do we
We never merge two sources' figures different methods over different periods are not addends
We never use a legal term in our own voice war crime, genocide, apartheid appear attributed, with the organisation, the date and the document — or not at all
We never promise completeness a broad inventory and a complete one are different things

What we have read, and what we have not

Four numbers, and they do not say the same thing. That is the point of giving four.

Source documents recorded 1 820, from 32 organisations
Read, with figures taken from them 1 512
Read, and the document carries no figure — established, not assumed 170
Read, but we received only a landing page: we conclude nothing 116
Out of reach by the means we allow ourselves 22

These five numbers are counted from the records at build time, not typed in. They were typed in once, and by the time anyone read them they were wrong.

⚠️ Why the third and fourth lines are separate. "This document contains no figure" is a statement about the document. "We only got a landing page" is a statement about us. A register that merged them would be reporting its own failures as the silence of its sources.

Why 22 documents are out of reach

All of them for the same reason: the publisher's site refuses our requests, and no public copy of the document exists anywhere else. Not one is refused because the publisher forbids us — that is a different thing.

⚠️ A site refusing our requests is not the same as a site forbidding us. Where a publisher simply blocks automated traffic, we read the same document in a public copy — the one its own record already pointed to — and the record says where we read it. Where a publisher names us in its rules, we do not go, and we do not go around. Not through an archive, not by any other door.

Nothing here is collected by a browser that presents itself as a person. Where a site refuses the request, the document is either absent from the register or, as with the Court above, obtained by hand and declared as such in the entry that uses it.

Where the register is uneven

The International Court of Justice no route we allow ourselves reaches its site. Its 13 documents were downloaded from an ordinary browser by this register's editor, and read here. No request was ever made to the Court, and every entry drawn from them says so
Israel's Ministry of Health does not answer us, while the Palestinian Ministry of Health does: 226 records on one side, none on the other
The Palestinian Central Bureau of Statistics does not answer us, while the Israeli one does
Hebrew and Arabic 463 sentences recorded in Hebrew, 276 in Arabic. Peace Now publishes a large Hebrew corpus that is also reachable; the Palestinian equivalents publish in English, or in files whose Arabic we cannot read in the right order

Each of these follows from what a source lets us reach, not from what we chose to record.

One has closed. B'Tselem — the Israeli documentation organisation — was unreachable when this page was first drafted. It is now read, through public archived copies, like everything else.


Where the figures come from, and why we always say

Most figures about this war are relayed, not established, by the organisation publishing them. A count can pass through three hands before reaching a reader who then attributes it to the last one.

So every entry names the whole chain. For example:

An academic report states that, as of August 2024, OCHA reported figures from the Palestinian Ministry of Health identifying more than 40,000 Palestinians killed.

Three links, three names. The origin is the health authority of one party to the conflict, and the entry says so.

Government sources, on all sides

Sources belonging to a party to the conflict are admissible and never neutral. They are entered, and they are marked. The Palestinian Ministry of Health and the Israeli State Comptroller carry the same mark, in the same words.


When sources disagree

We publish the disagreement instead of picking the heavier figure. Examples currently in the registry:


Corrections

A published entry is never left wrong.

Once a month we search every source we cite for a published correction, retraction or new version. When one is found, the entry carries a notice immediately — a notice states that a correction exists, which invents nothing. The finding itself does not change without a human reading the correction.

Corrections are logged with their date, their nature and the previous version, and the previous version stays visible. We do not correct silently.

The notice stays even when the correction turns out to change nothing.


Translations

Sources that do not publish in English are admissible. When we translate:

Charged words are rendered the same way on both sides. The Hebrew word an Israeli organisation uses for 7 October and the Arabic word a Palestinian ministry uses for an Israeli strike are both rendered as the source's own term, and both are marked as such.


One source excluded, and why

ACLED. Its licence forbids automated collection, republication and sub-licensing. Our public exports are CC BY 4.0, which is a sub-licence. This is not a technical obstacle: it is an incompatibility of substance, and it is the only exclusion of its kind in the registry.

Two publishers forbid us by name in their robots.txt — The Lancet and Springer. We never send them a request. We do cite their documents: citing is not crawling, and the address is there for a human reader.


What is not yet published

15 entries are published and 55 more are drafted and waiting. Entries go online at three to four a week, and no entry is published that a human has not read and approved.