// claim provenance, traced
Wikipedia gives you the claim. Origin Trace gives you its provenance.
Where did this
claim come from?
Origin Trace reconstructs the genealogy of a claim’s credibility (when it entered, whether it was born with a source, when the evidence changed), down to the exact revision. When the record doesn’t back it up, it says so, instead of inventing a provenance.
reads the revision historybisects to the originnever an LLM
“the world’s happiest animal”
- 2014claim introduced · no source
- 2019reworded · source added
- todayThe West Australian (2019)
sample-and-bisect · closed corpus
// the method
From a claim you read to the genealogy that holds it up.
Wikipedia stores the current article. It doesn’t store the story of how a claim earned its evidence.
The citation you see today says nothing about whether the claim was born with it, or whether a source was bolted on, years later, to launder an assertion that was never backed. That story lives only in the revision history. So Origin Trace reads the revision history.
Every verdict is reconstructed from the article’s revision history, read as wikitext and diffed revision to revision. Deterministic string work, not a model’s guess.
- 01
The claim
A sentence you can read in the article today.
- 02
Resolve the article
Which page carries it, or (when it’s ambiguous) the candidates to choose from.
- 03
Enumerate the history
List every revision, oldest to newest: finite and enumerable, a closed corpus. That closure is what makes silence provable.
1,690revisions · Quokka, enumerated
- 04
Locate the origin
Sample, then bisect (reading only what it needs) to the earliest occurrence it can confirm, and it flags when a sparse earlier one can't be ruled out.
- 05
Classify the evidence
Born-sourced, retrofit, unsourced: each verdict pinned to revision ids.
The hard case: a claim’s presence isn’t monotonic. It can be added, removed, and re-added across an article’s life.
A plain binary search would cave here, landing on whichever add-or-remove edge it happened to hit. So the search never trusts the first occurrence it finds: it re-scans earlier revisions for an older origin, and when the true birth may predate what it actually read (a re-addition, or a sparse island below the sampled range), it lowers its own confidence and says so, instead of pinning a false date.
// why the verdict holds
It’s not a fact-checker: it doesn’t say “true” or “false.” It says where a claim’s backing came from, and how solid it is, all auditable down to the exact revision.
01 · genealogy
Genealogy, not parallel
An LLM summarizes evidence by treating sources as independent. But belief doesn't spread in parallel. It spreads genetically. Origin Trace reconstructs the chain: who cited whom, and where the root is weak.
02 · closed corpus
Silence is provable
An article's revision history is finite and enumerable, a closed corpus. When the search reads every revision below an origin, “unsourced until 2019” is a proof about the whole history, not an “I didn't find it”. When it only samples that range, it says so rather than overclaiming. Abstention becomes trustworthy.
03 · honesty
It admits when the record is silent
When the wording shifts and the verdict depends on where you draw the line, the tool shows both readings and hands the judgment back to you, instead of inventing a certainty the history doesn't support.
// the vocabulary
It doesn’t say true or false. It classifies the evidence history.
Every claim resolves to one of these patterns: a read on the life of its evidence, not a fact-check of its content.
born-sourcedsourced
claim and citation entered together
retrofitback-filled
born unsourced; citation attached later
unsourced-stableunsourced
never sourced, never removed
ambiguousambiguous
the verdict depends on what counts as the claim
// cases
Cases.
A curated file of what the engine can catch: citogenesis, retrofit, unsourced-stable claims. Every verdict is real engine output, pinned and reproducible. Hit verify live on any of them.
// live
Interrogate a claim yourself.
Paste a claim. Origin Trace tries to resolve which article it belongs to, and when the scope is ambiguous, it shows you the candidates instead of guessing. Then it samples and bisects that article’s real history down to the revision that introduced the claim.
// audit
Or audit a whole article.
One read of the current revision maps every sentence to its evidence: which carry an inline citation, which assert without one. The claim boundary comes free from Wikipedia’s own structure. No NLP. Then click any uncited sentence to trace its history, down to the revision that introduced it, or an honest “not found” when the wording has drifted too far back to match.