StemReferences
Traceback. What does this message stand on? Every claim is standing on something: a source, a statistic, an earlier claim, or nothing at all. StemReferences maps the full support structure of a text, claim by claim, and follows each stem back to its root. Where the stem ends in solid ground, you get a citation graph. Where it ends in air, you get a flag, before your opponent, your regulator, or a reporter finds it first.
Use it to
- Pre-flight testimony, filings, and white papers where a single unsupported claim is a liability.
- Build a living evidence library: every claim your organization has ever made, mapped to its support.
- Reverse-engineer an opponent’s argument and find the load-bearing claim with the weakest stem.
- Satisfy the compliance officer who asks, correctly, “says who?”
Takes a text and the sources to trace against, and returns each claim with its support, a supported or unsupported status, and, where the stem ends in air, a risk rating and a note.
| Verb | Signature | Description |
|---|---|---|
trace | trace(text, sources) | Map a text’s support structure claim by claim and flag the stems that end in air |
graph = client.radiance.stemreferences.trace(
text=open("oped_draft.md").read(),
sources=["evidence_library", "web"],
)
for claim in graph.unsupported():
print(claim.text, claim.risk)That note field is the whole philosophy of the platform in miniature. We do not just find the hole. We tell you why it is the hole that matters, because we have stood at the podium when someone found ours.
ROI in one line: the cost of an unsupported claim is not the correction. It is every future claim being read with suspicion. StemReferences protects the asset your organization spent decades building: being believed.