ClaimTrace: Contradiction-Aware Evidence Lineage for Verifiable AI-Assisted Scientific Literature Reviews

ClaimTrace is a proposed framework for checking the relationship between scientific review assertions and their cited evidence. Its central mechanism is a typed lineage graph that preserves source identity, assertion boundaries, study conditions and numerical bindings rather than returning a single undifferentiated judgment. This paper specifies eight modules, a five-status annotation taxonomy, conservative contradiction rules and an audit record for evidence-grounded revision. A lightweight Python prototype was implemented and executed on twelve publisher-verified research abstracts. Four source documents were assigned to development and eight to evaluation; no source document occurred in both partitions. Seventy source-derived synthetic diagnostic cases were constructed, including forty-seven evaluation cases. The full prototype matched forty-two evaluation specifications, while a lexical retrieval-only heuristic matched nineteen. These counts measure conformance to constructed test expectations, not predictive accuracy against independent scientific annotations. Retrieval returned the generating source among the top three passages for all thirty-one evaluation queries with a known evidence target. The experiment did not execute a standard LLM judge, a learned natural language inference model, semantic retrieval, independent human annotation or authentic cross-study conflict evaluation. Consequently, the central question of whether typed lineage improves verification relative to retrieval plus an LLM judge remains unanswered. The contribution is a reproducible design and an executed, explicitly provisional implementation diagnostic, with the limitations needed for a subsequent controlled scientific evaluation.