``` ENTITY CHAIN BREAK AUDIT System/trace reviewed: Atlas Research Assistant — researcher education lookup Original question: What country is the university located in where the researcher who created the Atlas compression method earned her PhD? Hops in trace: 2 ── FINDING CARD 1 ── Location: Hop 2 handoff (Hop 1 → Hop 2) Entity: Dr. Maya Chen (resolved in Hop 1) Evidence: Hop 2 sub-question sent to the retriever: "Where did she earn her PhD?"; no name restated from Hop 1's answer ("Dr. Maya Chen"). Classification: RETRIEVAL-SIDE BREAK Reasoning: The unresolved pronoun "she" gave the retriever nothing to anchor on, and the retrieved chunks show the consequence directly: Chunk A is a generic, un-anchored statement about unnamed researchers, and Chunk B surfaces a different, merely similar-sounding person (Maya Ortiz, not Maya Chen). ── FINDING CARD 2 ── Location: Chunk B retrieved in Hop 2 Entity: Maya Ortiz (mistaken for Dr. Maya Chen) Evidence: Retrieved chunk: "Research scientist Maya Ortiz received her doctorate from Stanford University before joining the laboratory." This chunk names a different individual than the one resolved in Hop 1. Classification: NO BREAK Reasoning: There is no entity-chain break to classify here — the chunk is internally coherent and fully resolved on its own terms. It is simply the wrong entity, a symptom that traces back to the retrieval-side break in Card 1, not a separate generation-side failure of its own. ROOT-CAUSE VERDICT: RETRIEVAL-SIDE The entity resolved cleanly in Hop 1 but was never carried into the Hop 2 sub-question. That dropped name is what let the retriever return an unanchored generic chunk and a same-first-name false match, producing a final answer that names the wrong university with an appended (but insufficient) hedge. Refused inferences: none. The evidence in both chunks was sufficient to classify directly without guessing. Injection signals: none Source: arXiv:2609.06065 (Lee & Lim, UNIST, EMNLP 2026 Findings) ```