Verification Report 001
Emergence Verification Report 001
Method: Applied Verification, Dimension — Evidence Integrity.
Subject: Anthropic Claude Citations.
Claim under review
"Ground Claude's responses in your source documents. Citations return the exact passages that support each claim, so you can verify answers and surface sources to your users."
— Anthropic, Claude Platform Docs, "Citations"
This review was designed to test a public claim, not to confirm or discredit the vendor.
Finding
On a bounded adversarial test suite, the Claude citations feature returned only pointer-valid, support-valid citations. Every citation resolved to a verbatim passage in the supplied source document, and every cited passage substantively supported the claim it was attached to.
This finding is scoped to the tested cases. It is not a claim of universal reliability.
What we tested
The review asked one question: do Claude citations actually point to passages that support the claims Claude makes?
Two validity axes were scored independently:
- Pointer validity — does the citation resolve to a real passage in the supplied document?
- Support validity — does that passage substantively support the claim?
A valid pointer is not automatically supporting evidence; both were checked.
Six fixed documents and 27 prompts, covering: explicit facts, paraphrases, numerical derivations, multi-passage synthesis, ambiguous language, absent facts, contradicted premises, qualified claims, temporal and version distinctions, and multi-document source selection.
What we found
27 of 27 prompts produced evidence-backed answers.
- 0 invalid pointers — every citation was a verbatim substring of its source document.
- 0 non-supporting citations — every cited passage justified its claim.
- 0 fabricated citations.
- 0 fabricated answers on "absent fact" or "not established" questions.
Behaviors observed under pressure:
- Evidence and inference kept separate. When the model derived a number not stated in the source, it flagged the derivation as its own calculation rather than attributing it to the source.
- Section distinction. Given a document stating two different employee counts in two sections, the model identified both, explained the discrepancy, and resolved to the correct current figure.
- Version supersession. Given two policy versions where the newer superseded the older, the model cited the superseding document and noted the older figure was replaced.
- Correction handling. Given a document containing an explicit retraction (15% corrected to 8%), the model reported the corrected figure and flagged the retracted one.
- Refusal without fabrication. On questions the source did not answer, the model said so and cited the nearest relevant passage rather than inventing a number.
Limits of this finding
This is a bounded confirmation, not a reliability claim. The tests used one model (claude-opus-5), six documents, and 27 prompts. A larger randomized suite across models and edge-case documents would be required to support any claim beyond this scope.
Support validity was assessed by a human reviewer reading each cited passage against its claim; it is followable but not machine-scored.
The "guaranteed valid pointers" sub-claim was not stress-tested against malformed or extreme inputs (empty, very large, or unusual-format documents).
What would change this finding
A single counterexample — one pointer-invalid or non-supporting citation on a comparable prompt — would move the finding from PROTECTIVE toward CONCERNING. The full procedure is published so any such counterexample can be checked.
Evidence and reproducibility
Six fixed source documents were used, hashed at test time (SHA-256):
- Northwind Analytics 2024 Annual Report (simple prose)
- Meridian Health Q3 2025 Clinical Trial Update (structured, tabular, footnoted)
- Veridian Group Annual Report 2025 (long, multi-section, with internally inconsistent figures)
- Helios Institute Memo on Remote Work Productivity (adversarial: retraction, qualified claims, scope limits)
- Acme Data Retention Policy, Version 1 and Version 2 (multi-document supersession)
27 prompts across ten categories: explicit fact, paraphrase, numerical derivation, multi-passage synthesis, ambiguous language, absent fact, contradicted premise, qualified claim, temporal/version distinction, and multi-document source selection.
- Model: claude-opus-5, Anthropic Messages API, citations enabled.
- Raw API responses: preserved in full (text and citation objects) for all 27 prompts.
- Document hashes: SHA-256 recorded per document.
- Procedure: supplied. Any reviewer with API access and the same documents and prompts can reproduce the run or challenge any finding.