About WikiDrift

Measures how a Wikipedia article changed over time, and how language editions of the same topic differ. Public data, revision IDs, reproducible diffs. Part of the encyclopediae.org family — the diagnostic arm beside the constructive project of encyclopedias from academic institutions. Open source on GitHub.

Candidates only — not conclusions.

What it measures

WikiDrift runs on public MediaWiki and WikiWho data. For each article it can report:

• History: when long-stable text was dismantled (pivots), with before/after diffs and which accounts removed or wrote the text (persistence-weighted displacement).
• Other editions: how English compares with editions such as Hebrew, Arabic, Polish, or German on framing and load-bearing facts — including whether English moved at the pivot, not only whether editions differ today.
• Citations: how the article's own source mix shifted across a rewrite (from → to), without reliability ratings on domains.

How findings are produced

Articles are selected and scored from content trajectory, not from a roster of suspect editors. Optional named lists can sit on top as overlays; they are not the foundation. A metadata pre-ranker (byte displacement, no full text) decides what deserves a full token-level pass and routes addition- or churn-heavy cases to framing analysis when removal alone would miss them.

A large rewrite is a change signal. Base-rate runs show legitimate overhauls can out-pivot contested topics. Direction is a separate layer (stance + cross-edition checks). Every page here is a lead with links back to underlying revisions.

Limits in brief

Internal history cannot see bias that was there from the start with no rewrite to contrast (born-framed / long-standing consensus). Those cases need external reference — other language editions and related instruments. Attribution reports public actions only.

How articles were selected

The articles on this site are a validation set, not a curated accusation list. The thesis cluster (Israel-Palestine, Holocaust in Poland) was chosen because those articles appear in independent external sources — Wikipedia's own ArbCom arbitration findings, a peer-reviewed academic paper (Grabowski & Klein 2023), and a 2025 ADL report. Those sources define what the detector should find; the detector runs on each article's own content independently and must reach the same conclusions without consuming any list.

Clean controls (Photosynthesis, Brontosaurus) and cross-domain contested articles (Climate change, Abortion) were added to establish the base rate and catch false positives. The tool works on any article.

How to use it

Open a finding, read the tabs, follow the revisions. For layers, validation lessons, principles, and citations, see Methodology.