Open Data

Take the data
and check it

Most of what circulates in this argument is an assertion with no numbers attached, or a number with no source attached. Everything below is the opposite: the raw data behind this site's original research, free to download, reuse and republish, including by people who want to prove us wrong.

These are not collections of other people's statistics. They are original comparative work that exists nowhere else: a conflict-by-conflict comparison built to a single consistent measure, and an article-by-article audit of how one encyclopedia frames one subject.

Licence: CC BY 4.0. Use it commercially, modify it, build on it. The only condition is attribution. If a dataset here changes your conclusion, publish that too.

42 records · JSON + CSV · version bd2270ae

Conflict comparison: deaths, rates, and international response

Comparative figures for the Gaza war set against other conflicts of the same era, across five measures: total dead, dead per year, dead within a matching time window, UN Human Rights Council condemnations, and elapsed time before an international court acted. Each row names the perpetrator. Where a state obstructed counting, a documented floor and a higher estimate are both recorded.

Shoresh Wikipedia audit, dataset conflict-comparison, version bd2270ae.
https://ilroot.com/data/conflict-comparison.json
sha256 bd2270ae3aa8987b4f07c2ea997c15452fb7c3ccf90cb7f9af794e7e7dc053cc · CC BY 4.0
992 records · JSON + CSV · version 702ad2fa

Wikipedia bias audit: Israel-related articles

An audit of English Wikipedia's Israel-related coverage, cataloguing 859 articles against 23 recurring framing techniques, alongside 133 articles checked and found clean. Every entry is pinned to the exact Wikipedia revision that was live when it was audited, so each finding stays verifiable even after the article changes, and the dataset records which articles have been edited since. Checked 2026-06.

Shoresh Wikipedia audit, dataset wikipedia-audit, version 702ad2fa.
https://ilroot.com/data/wikipedia-audit.json
sha256 702ad2fae5615ce23753ea80b596e2fe37b91c0a49da8c34d814e15a5160eff7 · CC BY 4.0
72 records · JSON + CSV · version 28af8849

Fact-check corpus: claims about Israel and the Jewish people

72 circulated claims, each with a verdict, a sourced summary, and a permanent URL. Ratings use a 1 to 5 scale where higher is closer to true: 1 false, 2 misleading, 3 contested. Every entry carries schema.org ClaimReview markup on its own page.

Shoresh Wikipedia audit, dataset fact-checks, version 28af8849.
https://ilroot.com/data/fact-checks.json
sha256 28af8849a6540ef1e10330e75feacc8a262bc4fdcdf5c9dbdac9d6e903cb131a · CC BY 4.0
28 records · JSON + CSV · version 2c809f5e

Primary texts of the conflict, complete and checksummed

28 founding documents, 80,858 words of primary text: declarations, mandates, UN resolutions, the 1949 armistice agreements, charters and peace treaties. Each was transcribed once from a named scholarly repository, is served complete on this site, and carries a SHA-256 of its exact wording so any copy can be verified against ours. The full text of every document is available in one file as primary-texts-corpus.json.

Shoresh Wikipedia audit, dataset primary-texts, version 2c809f5e.
https://ilroot.com/data/primary-texts.json
sha256 2c809f5e8bd155b4061a7e1d833f67defbd56fcb970c94dc658fceb2283ce67c · CC BY 4.0
1,306 records · JSON + CSV · version 18d50695

Source archive: every citation, and whether it still answers

Every outside page this site cites, 1306 in all, with its publisher, whether the original is still reachable, and a link to the copy we hold. We keep the readable text of 658 of them, captured on the date recorded and fixed by a SHA-256, so a citation survives the page being moved, paywalled or deleted. Wikipedia citations are not copied; they are pinned to permanent revisions instead. The original link is always the primary source and our copy is only ever a fallback.

Shoresh Wikipedia audit, dataset source-archive, version 18d50695.
https://ilroot.com/data/source-archive.json
sha256 18d506950780718997f70c9d3ac3435c9d781edeae8635815b43bc19cfea8f7c · CC BY 4.0
12 records · JSON + CSV · version 6272a88d

Correction log

Every correction made to this site: what it said, what it says now, why it changed, and whether the error was reported or found in an internal audit.

Shoresh Wikipedia audit, dataset corrections, version 6272a88d.
https://ilroot.com/data/corrections.json
sha256 6272a88d84b3bd5aa1c24cd3ef728f4b789cfd36b4afcff3026438d22722a990 · CC BY 4.0
4,921 records · JSON + CSV · version 2cddf626

Wikipedia: completed erasures

Article openings that carried a tracked word since 2005 and carry it no longer, with the verdict checked against the full current article.

Shoresh Wikipedia audit, dataset wikipedia-erasures, version 2cddf626.
https://ilroot.com/data/wikipedia-erasures.json
sha256 2cddf626a9554abbf624632ecca78e75817fb48ab0734c723d66157264d5e990 · CC BY 4.0
278 records · JSON + CSV · version c41fceaf

Wikipedia: pinned removals

Each removal traced to the exact Wikipedia edit that made it, with the public attribution Wikipedia itself displays.

Shoresh Wikipedia audit, dataset wikipedia-pinned-edits, version c41fceaf.
https://ilroot.com/data/wikipedia-pinned-edits.json
sha256 c41fceaf4d926c6ac00a0d060f71d0805ac24cd1c782e55b14b18f6e9e8d3111 · CC BY 4.0
195 records · JSON + CSV · version fd829dd5

Wikipedia: merged and renamed articles

Watched articles that no longer exist under the name we recorded, resolved to their current target.

Shoresh Wikipedia audit, dataset wikipedia-renames, version fd829dd5.
https://ilroot.com/data/wikipedia-renames.json
sha256 fd829dd5b1099e3b061f4520d0c92856b3aba1836aa1b301be9442ec67335343 · CC BY 4.0
1,000 records · JSON + CSV · version 37b4ffa0

Wikipedia: offender index

All watched articles scored on six signals (severity = zone reach x magnitude; article score = strongest finding + a diminishing sum of the rest, x (1 + prominence)); top 1000 published.

Shoresh Wikipedia audit, dataset wikipedia-offender-index, version 37b4ffa0.
https://ilroot.com/data/wikipedia-offender-index.json
sha256 37b4ffa04362b11f49fcd9e4fe74e93e3af2e1f263ab5af21542ba2d6fed4cad · CC BY 4.0
1,569 records · JSON + CSV · version 4ed6e628

Wikipedia asymmetry test: designation and accusation vocabulary by perpetrator

1569 articles classified by which side the encyclopedia's own categories say perpetrated the event: 1399 palestinian_perpetrated, 170 israeli_perpetrated. Each row carries the matched category, the first and last waypoint held with its date and revision id, and the designation and accusation word counts (DESIG_TERMS, ACC_TERMS, both printed on /data) at each end. Recompute the rate yourself: the share of each bucket where acc_last exceeds acc_first is the number this site's headline ratio is built from.

Shoresh Wikipedia audit, dataset wikipedia-asymmetry, version 4ed6e628.
https://ilroot.com/data/wikipedia-asymmetry.json
sha256 4ed6e628c8026b9c0d87574ca22630e03220a9943d0229026ebdbba57a381748 · CC BY 4.0

How to cite

Any of these forms is fine. The point is that a reader can get from your citation back to the row. Every dataset card above also carries its own versioned form, with a content hash so a citation names the exact data behind it, not just the page.

Shoresh (ilRoot). "Wikipedia bias audit: Israel-related articles." ilroot.com/data. CC BY 4.0.

Method, and its limits

Stated plainly, because a dataset whose method you cannot inspect is just an assertion in a spreadsheet.

The conflict comparison uses each conflict's most widely accepted total, and where a state obstructed counting it records both the documented floor and the higher estimate rather than splitting the difference. For Gaza it uses the Hamas-run health ministry's own maximum figure, unadjusted, so the comparison survives being handed the other side's best number. Rates are simple arithmetic: total divided by elapsed years, which flatters no one and is easy to check.

The Wikipedia audit quotes the live article wording on the date checked and links to it. Wikipedia changes, so an entry is a claim about a specific article at a specific time, which is why the check date travels with the data and why the set includes articles examined and found clean. A finding is not "this article is biased"; it is "this specific sentence does this specific thing", and you can open the link and disagree.

The asymmetry test, behind the wikipedia-asymmetry dataset, classifies an article by what the encyclopedia's own categories say happened, not by our reading of the prose: an article whose categories match the first pattern below is filed palestinian_perpetrated, one matching the second is filed israeli_perpetrated, and an article matching both or neither is dropped rather than guessed at. Each row then counts, at the first and last waypoint held, how many words from each list below sit in the opening. Reproduced exactly as the code runs them, case-insensitive:

PAL_PERP = /palestinian terroris|palestinian political violence|hamas|islamic jihad|pflp|fatah|suicide bombing|terrorism in israel|terrorist incidents in israel|massacres of israelis|attacks on israeli|murder of israeli|kidnapp\w* of israeli|rocket attacks|hezbollah attacks|arab riots|antisemitic attacks/i ISR_PERP = /israeli war crimes|massacres committed by israel|israeli military operations|israel\w* attacks on|killed by israeli|israeli airstrikes|idf operations|massacres perpetrated by israel|israeli occupation/i DESIG_TERMS = ["terror", "militant", "islamist", "jihad", "massacre", "antisemitism", "martyr", "hostage"] ACC_TERMS = ["genocide", "apartheid", "occupation", "colonial", "settler_colonial", "ethnic_cleansing", "war_crime", "famine", "siege", "displacement"]

What none of this is: a random sample. These articles and conflicts were selected because they bear on one argument. That is a real limitation and it is the reason the controls are published alongside the findings rather than quietly dropped.

Publishing the raw data is the part most advocacy never does, because raw data is the thing that can be turned against you. That is exactly why it is here. A claim you cannot audit is worth nothing, and we ask readers to audit everyone else.

If you find an error in any of it, report it. It goes in the public log with the old value, the new value, and who caught it.