Track changes to a Wikipedia article over time

Goal

Use Refract to analyze the revision history of a Wikipedia page and understand what changed, when, and what kind of change it was.

The analysis is deterministic: the same revision range produces the same events.

Steps

1. Run your first analysis

Zero install:

npx @refract-org/cli analyze "Earth" --depth brief --json

Or with a local install:

bun add @refract-org/cli
refract analyze "Earth" --depth brief --json

--depth brief reads the latest 20 revisions and leaves the changed text out of each event. detailed, the default, includes the text, and without a range reads the whole history: tens of thousands of revisions for this page. Step 3 shows how to give a range.

The -c flag caches revisions in SQLite. It needs Bun and a source checkout, and a cached range is reused only when both --from and --to are given.

2. Read the output

With --json, refract analyze prints one JSON event per line to stdout. Without it the command prints a text summary, or opens the explorer in an interactive terminal. Each event represents a single observed change at a revision boundary. For example (text values illustrative, enrichment fields left out):

{
  "eventType": "section_reorganized",
  "fromRevisionId": 1250019297,
  "toRevisionId": 1251058746,
  "section": "Geology",
  "before": "Earth's crust consists of tectonic plates...",
  "after": "Earth's crust consists of several tectonic plates...",
  "deterministicFacts": [
    { "fact": "section_changed", "detail": "change=modified" }
  ],
  "layer": "observed",
  "timestamp": "2024-10-14T00:00:00Z",
  "schemaVersion": "0.5.0"
}

Key fields to read:

eventType: what kind of change (sentence, citation, template, revert, section, etc.)
section: where in the page the change happened
before / after: the text before and after the change
deterministicFacts: why the engine produced this event — always mechanical
layer: "observed" means directly extracted from the diff; "policy_coded" marks events matched against a rule, such as revert_detected; "model_interpretation" is set by downstream consumers

3. Scope to a specific revision range

Large pages can produce thousands of events. Narrow the range with --from and --to:

refract analyze "Earth" --from 1250019297 --to 1258622167 --depth detailed --json

That range is 23 revisions from late 2024. Or filter by time with --since, which reads every revision from that moment to the newest:

refract analyze "Earth" --since 2024-01-01 --depth detailed --json

4. View results in a browser

refract explore "Earth" --from 1250019297 --to 1258622167

Opens a local web server (default port 8899) with an interactive timeline, evidence table, and diff viewer. Press Ctrl+C to stop.

5. Export structured output

# Line-delimited JSON for tools
refract export "Earth" --from 1250019297 --to 1258622167 --format ndjson > earth-events.jsonl

# CSV for spreadsheets
refract export "Earth" --from 1250019297 --to 1258622167 --format csv > earth-events.csv

# Bundle with a SHA-256 hash of its contents
refract export "Earth" --from 1250019297 --to 1258622167 --bundle > earth-bundle.json

export runs at detailed depth and, without a range, reads the whole history.

6. Track a specific claim

refract claim "Earth" --text "Earth is the third planet from the Sun"

This shows the matching sentence in each of the latest 50 revisions, and when the passage was first matched, changed, removed or reintroduced within them. --limit reads more revisions; --from and --to select an earlier range.

Understanding what you see

Citation events (citation_added, citation_removed, citation_replaced) trace source churn. A section with frequent citation turnover may indicate an actively contested topic. A citation that survives many revisions without change is a high-stability source.

Revert events (revert_detected) mark edits that were undone — either by the same editor ("self-revert") or by another editor. Clusters of revert events in a short time window often indicate edit-warring.

Template events (template_added, template_removed) reveal policy tagging. Templates like {{citation needed}} or {{NPOV}} are Wikipedia's dispute signals. Refract detects these mechanically from wikitext.

Section events (section_reorganized, lead_promotion) track structural reorganization. When material moves from body to lead, the framing of the page has changed — Refract flags this as a change in editorial emphasis.

Troubleshooting

Rate limits: Wikipedia's API limits request frequency. Keep runs to a revision range; Refract retries a 429 after the server's Retry-After.
Large pages: Pages with thousands of revisions produce many events. Use --from / --to or --since to scope to a date range.

Other wikis: Point --api to any MediaWiki instance:

refract analyze "Erde" --api https://de.wikipedia.org/w/api.php --depth brief --json

Too much output: Pipe to a file instead of stdout, or use refract export for structured formats.