Describe the requirements for sources and the verification plan. Return JSON {"question":"...","source_requirements":"...","check_plan":"..."}. question—research question; source_requirements - what sources are needed (type, freshness, how much, what to indicate next to each fact); check_plan - how will you check that the sources are real and confirm the fact (2-3 specific steps).
Mechanics: the model can sound convincing without relying on facts - it completes a plausible text. A link turns a statement into a verifiable one: if there is a source next to the fact, you can open it and check it. Therefore, sources are not decoration, but a way to distinguish knowledge from fiction. Levers. Type: ask for primary sources (official data, documentation), and not retellings of retellings. Freshness: “sources not older than 2 years.” Anchor: require a link next to each key fact, rather than a general list at the end - this way you can see what is supported by what. Inside trick: add the rule “if there is no source, write so, don’t make it up.” This dramatically reduces confident lying: the model begins to flag weaknesses instead of masking them. The second technique when checking: check that the source exists at all and says exactly what it says. Quick test - ask for an exact quote from the source; the fictitious link on the quote breaks. A common mistake: taking a nice list of links as evidence. The link may be real, but the conclusion from it may be fake. Check the fact versus source relationship.
Unlock access to submit solutions for instant AI review.