verified-research

Search however you judge best: breadth, depth, and query choice are yours. This skill governs only what reaches the user: every claim carries a source the user can audit, and every quote is checked by a script, never by your own say-so.

verify_quotes.py sits next to this file. Run it as python -X utf8 verify_quotes.py .... If it asks for an optional extra (pypdf for PDFs, playwright for JavaScript pages), install it once and rerun.

Steps

1. Collect. Search widely and in parallel. Prefer the primary source (official docs, release notes, government pages, the paper itself) over anyone describing it. Done when each claim you plan to make has a candidate source page.
2. Quote from the raw page. Read each source with verify_quotes.py --text URL and copy a sentence verbatim. Fetch tools that summarise the page paraphrase it, and a paraphrase fails the check. Done when every claim has a verbatim quote. When a site refuses automated access, respect it: cite a copy meant for machines (official API, DOI or abstract page, RSS feed, archive) or list the claim as unverified, never a quote from memory.
3. Weigh sources by the rules below. Done when every claim rests on a primary source or on two independent sources.
4. Write the answer in the user's language: conclusion first, each claim footed [^N], then a short Unverified section with that heading in the user's language (claims you could not source, sources that disagree and which one you trust and why). Everything above Unverified is footed, and its links live in footnotes (the script fails a link it cannot check). Footnotes, one per line, exactly: [^1]: https://example.com/page "verbatim quote" Join omitted parts of a quote with …. Each quote must carry the whole claim it foots; split a claim rather than stretch a quote over it. Omit the Unverified section when it would be empty.
5. Verify. Save the answer to a temp file and run verify_quotes.py <file>. Fix or drop every failing quote and rerun until it exits 0 (a fetch failure may be transient: rerun once before dropping). Unfetchable pages stay only if listed under Unverified.
6. Cold read, when the conclusion is a judgment, comparison, or recommendation, or when sources conflict. A plain fact settled by a verified primary-source quote is done at step 5. Hand the answer (not your notes) to a fresh subagent: "For each footnote, read the URL with python -X utf8 <path>/verify_quotes.py --text URL (summarising fetch tools drop text) and judge whether the quote actually supports the claim it is attached to. Then search for evidence against the conclusion. Report mismatches and counter-evidence." Fix what it finds and rerun step 5. Without subagents, do this as a separate pass that reads only the answer.

Weighing sources

• Primary beats count. One official page outweighs any number of articles repeating it; when they conflict, the primary source wins and the conflict goes under Unverified.
• Independent agreement only. Sources that copy each other (same wording, same numbers, citing the same origin) count as one.
• How the page was made. Weight follows the work behind it: measurements, tests, method, or raw data from an author who can be held to it (individuals and video creators included) carry weight. Plausible assertion, rankings without method, and mass-produced SEO or AI-written pages carry little even when not wrong; use them only to find what they point at.
• False premise. If the question assumes something that does not exist (an "official recommendation" no official page states), answer that it was not found rather than supplying the nearest number. Likewise when the primary source exists but omits the detail asked (a date, a figure): say the primary source does not state it, then give the best secondary evidence labelled as such.