Make your agent label every claim by source and how well it's confirmed

How it works

1. In your agent's standing instructions, rank source types from highest to lowest. First come primary documents, such as company filings and releases, official calendars and exchange notices. Next are call transcripts and interviews, then news and commentary. Social posts come last.
2. Your own notes rank highest for your own rules and intentions, but they never count as proof of an outside fact.
3. Define five labels in the same words you give the agent. "Confirmed" means it matches a linked primary document. "Direction only" means a primary document confirms the direction but not the size or exact figure. "Partly confirmed" means a linked transcript or news article supports it but no primary document does. "Disputed" means linked sources disagree. "Unconfirmed" means no link, or the only support is a social post or memory. A claim supported only by a social post is unconfirmed.
4. Require every factual line in a summary, prep note or research page to carry its source link and its label.
5. When two sources disagree, the higher-ranked one is used and the conflict is written down. The agent never averages or blends the two.
6. Unconfirmed items go in a separate list at the end and are never used in the summary's conclusions.
7. Once a week, open three random "confirmed" links and check them yourself.
8. Log the weekly check in your journal. Open the journal, pick three "confirmed" items at random (a short script, a dice roll, or a shuffle), open each original source, verify the claim matches, and note any discrepancies you find.

Copy-paste prompt

Standing rules for every summary or note you write for me. You do not give trading advice.

Source ranking, highest first:
1. Primary documents: company filings and releases, official economic calendars, exchange notices.
2. Transcripts of calls and interviews.
3. News articles and commentary.
4. Social media posts.
My own notes in {{notes_folder}} define my rules and intentions. They are not evidence for outside facts.

Labels, exactly one per factual line: confirmed = matches a linked primary document; direction only = a primary document confirms the direction but not the size or exact figure; partly confirmed = a linked transcript or news article supports it, but no primary document does; disputed = linked sources disagree (still cite the higher-ranked source and add a Conflict line with both links; never average or blend); unconfirmed = no link, or the only support is a social post or memory. A social post alone is never confirmed, direction only, or partly confirmed.

For every factual line: add the source link and one label. No link means the label is unconfirmed.
If sources conflict, use the higher-ranked one and add "Conflict:" with both links.
Two articles repeating the same claim count as one source unless each cites a different primary document.
Put all unconfirmed items in a final "Unconfirmed" list and leave them out of any conclusion.
If you don't know, write "unknown". Never fill a gap from memory.

Guardrails

• No link, no "confirmed". A label is only upgraded when a new, higher-ranked source is added.
• Copies of the same story count as one source.
• Unconfirmed items never appear in conclusions.
• The agent writes "unknown" rather than guessing, especially for dates and figures.
• Labels must use footnotes or a compact format so output stays readable; never repeat the full link on every line.

What can go wrong

Without a ranking, models treat a confident news article and a company filing as equals and quietly blend them into one number. The other failure is label inflation: the agent marks a claim "confirmed" because it found it twice, both times in articles copying the same source. Defining "confirmed" as "matches a primary document" fixes most of it. Expect the model to ignore or creatively apply the ranking and labels anyway: marking news as "confirmed", inventing plausible-looking links, or mislabelling under its own confidence. The weekly spot-check of three items is a sample, not a net; it will not reliably catch this at scale, so treat every label as a prompt to verify, not a guarantee.

Source and references