Methodology

Rows

Each row is one death from the OCHA oPt casualties dashboard: Palestinian fatalities, area West Bank, perpetrator Israeli civilian settler, read month by month. The dashboard has no names, so a row is a month and a place.

Every run first re-reads 2023 and compares it with a locked copy. If the dashboard is offline, blank or different, nothing is published.

OCHA's sentences

For each row we read every OCHA report from the first day of the death's month to 60 days after it ends: OCHA's Protection of Civilians reports, then Flash Updates. A model proposes which sentences describe the death. Code then checks each one: it must appear word for word in the stored report, be one sentence or part of one, and have 40 words or fewer. A row that fails is held and listed.

Running totals (for example, the number of deaths since January) never tag a row. One such count of attackers is shown on its own.

Tags

A tag needs a sentence about this death that contains one of these phrases:

• Attacker claim: "stabbed" or "stabbing", "attacker", "attempted ... attack", "perpetrator"; or "reportedly", "allegedly" or "according to the Israeli military" together with a weapon or attack word. The last group shows as "alleged attacker".
• Perpetrator doubt: "by Israeli forces", "by a soldier", "settlers or forces".
• Two-sided clash: "threw stones at each other", or Palestinians and settlers both throwing stones or firing.

When the model reads a tag in a sentence that uses none of these phrases, the row is held until a person reviews it. Each decision (accept, reject, or add a tag) is recorded with the exact sentence, and the row shows "tag set by human review".

A sentence that states a number of deaths (for example "four were killed") is attached to at most that many rows.

Israeli-side accounts

For each death we add, where found, a statement by the Israeli army, police or Shin Bet as reported by an Israeli news outlet, and lines from audits of OCHA's data by Regavim and the Kohelet Policy Forum. Each one passes the same word-for-word check against the stored page. They sit in their own column and never change A, B or C. When two outlets carry the same statement, it counts once.

A, B and C

• A: deaths the dashboard shows.
• B: A minus rows with an attacker claim in OCHA's text.
• C: B minus rows with doubt about the perpetrator in OCHA's text. C still includes two-sided clashes.

The three numbers show what each counting rule gives. None of them is an estimate of the true number, and none is a count of any legal status.

Use of OCHA material

OCHA's content is quoted with credit and a link, following the UN's terms for news-related material. The site does not offer a download of OCHA rows.

Sources and credits

• UN OCHA oPt: the casualties dashboard and the reports every row and quote comes from (ochaopt.org).
• Regavim: its audit of OCHA's settler-violence data, the source of the audit lines (public sheet).
• Kohelet Policy Forum: audit lines, quoted with the report they come from.
• Israeli and Palestinian news outlets: each line links to its article.
• Map: Oslo areas and firing zones from OCHA, settlements from Peace Now, places from OpenStreetMap contributors. The Area C building count on the map page is Regavim's, quoted from its report.

We wrote to OCHA and to Regavim before the public launch.