A git log for Wikipedia
Refract shows how a page's sentences, citations, and structure changed, revision by revision, with the source revisions behind each observation. It works on Wikipedia and any other MediaWiki wiki.
It runs locally under deterministic rules, with no API key or model. Your review process decides what the changes mean.
Start with one change
Section titled “Start with one change”On 24 August 2006, the International Astronomical Union reclassified Pluto as a dwarf planet. Wikipedia's Pluto article went through 43 revisions in under an hour that afternoon. This prints its first sentence from each of them:
npx @refract-org/cli@0.6.1 claim "Pluto" --text "ninth and smallest" --from 71593043 --to 71601937The sentence starts at "Pluto is the ninth…" and passes through "Pluto was until recently considered the ninth…". The captured walkthrough links each change to its revisions.
For repeated review, review a watchlist. Refract remembers the last revision you explicitly acknowledged.
What you receive
Section titled “What you receive”Changed text
Sentence observations include the wording before and after the transition.
Source changes
Citation events identify references added, removed, or replaced.
Revision context
Events carry revision identifiers, sections, timestamps, and analyzer provenance.
Repeatable observations
The same captured inputs, analyzer versions, and configuration produce the same observations. Export metadata can vary.
Files for your tools
Export NDJSON for a notebook, SQL query, browser viewer, or downstream review process.
Inspectable history
Follow the analyzed revision span. Hash verification checks record consistency; it does not establish complete coverage or authenticate a source.
Choose a workflow
Section titled “Choose a workflow”Review sources you revisit
Inspect pending changes since your last acknowledgment. Keep fetched progress separate from reviewed progress.
Inspect an earlier record
Find a revision at or before a date, or compare a bounded span of revisions.
Use changes as data
Select events such as removed citations and inspect their source context in your own tools.
Available revision histories
The CLI fetches Wikipedia and compatible MediaWiki histories. Library consumers can supply versioned text through the shared revision model.
Observation limits
An arbitrary web URL does not supply a revision history. First observed means first in the analyzed range. Wording changes alone do not establish a dispute or whether a statement is supported.
One repository, optional interfaces
Section titled “One repository, optional interfaces”| Interface | Use it for |
|---|---|
| CLI and libraries | Fetch, compare, review, and export source histories; use the built-in explorer |
| Browser viewer | Inspect an exported NDJSON event file locally |
| Python adapter | Call the CLI from notebooks and read events as dataclasses or DataFrames |
All three live in the Refract repository. Event contracts and reproducible examples are maintained with the engine. The public review and history-range pages are built directly from those core guides.
License
Section titled “License”The engine is AGPL-3.0. This site's original documentation is CC-BY-4.0; included core guides retain their source license. Built and maintained by NextConsensus.