| name | claim-verification |
| description | Use when a debate agent needs to judge whether a candidate's claim is supported, contradicted, or unaddressed by retrieved evidence from the professor's own publications. |
Claim Verification
Judge ONE candidate claim against a set of retrieved evidence chunks pulled from
the professor's own indexed publications.
Procedure
- Read the claim in isolation — do not assume context beyond what it states.
- Read every retrieved chunk. A chunk supports the claim only if it explicitly
and specifically overlaps with what the claim asserts — topical adjacency is
not support (e.g. "the professor works on transformers" does not support
"I built a chatbot" merely because both are ML).
- Decide one of:
- "verified": at least one chunk directly and specifically supports the claim.
- "refuted": the retrieved evidence directly contradicts the claim.
- "unclear": no retrieved evidence bears on the claim either way.
- When verified or refuted, cite the exact chunk text you relied on as the
receipt. Never fabricate a receipt — if you cannot point to a specific chunk,
the verdict must be "unclear".
- Be conservative: a candidate's claim about their OWN work (e.g. "I published
at NeurIPS") is not verified by finding the professor's own unrelated paper —
only by evidence that actually speaks to the candidate's claim.