Glossary

Terms used on article pages and in the methodology.

Lead: A candidate worth inspecting. Not proof of bias or bad faith.
Drift: Narrative or content shift over successive edits.
Pivot: A window where a large share of long-standing text was replaced at once. Discovered from content displacement (not only from edit bursts or a known tag date). Shown on the Diff tab timeline.
Stable-then-retrofit: Long-surviving text dismantled and the replacement sticks. Structural signal for L1; still not proof of bias.
Framing: How a subject is portrayed (critical / neutral / sympathetic) beyond bare facts. Editions can frame the same topic differently.
Stance / NPOV axis: Framing scored critical–neutral–sympathetic. Not positive / negative sentiment (tone ≠ viewpoint balance).
Cross-lingual comparison: Same topic across language editions. Can be static (today) or pivot-relative (before vs after a detected rewrite). Disagreement is a signal with more than one possible cause.
Fact divergence: Incompatible load-bearing claims across editions (e.g. different counts). Separate from framing: facts can agree while frames differ.
Citation-source change: How an article's own citations shifted from → to across a major rewrite: domains added/dropped and mix of journal / news / book / web. No reliability labels on sources — composition only.
Blame: Per-span authorship: which account introduced which text (like VCS blame).
Editor concentration: Share of current text written by a small set of accounts (e.g. top-10), plus distinct-editor count. Context only.
Born-framed: Strong framing from creation, so there is no earlier baseline to compare. Needs an external reference, not a change detector alone.
Reframe-by-churn: Article net-grows while shedding unusual amounts of older text relative to its own baseline. May look HEALTHY to pure removal-ratio L1; routed to L2 via pre-rank.
Control: A neutral topic (e.g. Photosynthesis) used as a sanity check.
Conflict weight / M-score: How hard the article was edit-warred (mutual-revert style measures). Contested ≠ biased. Low score on a large change means the rewrite was not fought over.
Revisions: Exact revision IDs and dates behind a finding.
Stability prior: Long-surviving text is treated as sticky by default; a lasting collapse of that text is the anomaly the L1 layer targets.
Persistence-weighted loss (PWR): Rewrite magnitude: each token weighted by how long it had survived. Deleting old text counts more than churning recent text. Based on published content-survival metrics (Halfaker; Adler–de Alfaro). Ratio classifies; absolute mass ranks.
Change vs origin: By change: rewritten over time (this tool's L1 core). By origin: framed from the start or held steady without a rewrite — needs an external reference.
Base rate: Large rewrites are common and often legitimate. One rewrite in isolation is not diagnostic; compare against controls and other layers.
External reference: Something outside the article's own history — other language editions, other encyclopedias — used when internal history has no contrast.
Conjunction: Multiple stacked signals (stable → removed → meaning shift → persisted → concentrated authorship). Single factors are weak.
L4 discovery: Use destroyers of a confirmed pivot only as a search prior; re-test each candidate article on its own content. Graph membership never flags an article.