Source Gathering

Source Gathering

Retrieve the evidence. Do not interpret it — that is a different agent's job, and the separation is what stops a search-result snippet being reported as a source that was read.

When to use

• A questions row needs evidence it does not have.
• A finding rests on something nobody has actually retrieved.

Steps

1. Read sources first. The commonest waste here is fetching what is already downloaded, and the second commonest is re-fetching something already discarded as unusable.
2. Search to discover, fetch to retrieve. web_search finds a locator; web_fetch retrieves what is behind it. Nothing in the web namespace can find a URL on its own, and a snippet is not a source.
3. Prefer the original. A paper over a press release about the paper, a filing over an article about the filing, a dataset over a chart drawn from it.
4. Record each one with record_entry on sources: the locator, and kind as primary, secondary or commentary. Be honest about kind; it is what most evidential disagreements turn out to be about.
5. Note the interest. Who produced this and what they gain from it, in a sentence. Not a score.
6. Stop at gathered. Do not summarize what you retrieved. establishes is filled in by whoever reads it.

Output

sources rows in gathered, each with a locator somebody else can re-open, and nothing said about what they mean.