ExpertTwins: Auditable Long Horizon Scientific Panels through Grounded Expertise

Abstract

Multi-agent language-model systems increasingly assign roles to agents and let them deliberate. A role label, however, does not specify which evidence an agent can bring or which standard it applies before accepting a conclusion. Agents that read one another's outputs tend to converge. We present ExpertTwins, a framework in which each seat of a scientific panel is defined by an inspectable evidence policy. Discipline seats retrieve under field-specific queries, and scientist-associated seats retrieve from attributed publications and a citation-derived scholarly neighborhood. Private initial packets, structured exchange, cited-source disclosure, deterministic quotation checks, and peer source review turn deliberation into an auditable record, which we also render as interactive idea-flow maps. Scientist associations denote publication-grounded perspectives, not replicas or endorsements of named researchers. In a nine-seat, four-turn neuroblastoma panel, 438 of 439 claims passed all attached quotation checks, and seats withdrew or defended positions when challenged on their evidence. A ten-turn panel sustained evidence acquisition and differentiated recommendations without conversational collapse. Across six base models given byte-identical initial packets, four sustained 97.9-99.8% quotation grounding, and at least 91% on every turn, whereas Haiku 4.5 failed by paraphrase and GPT-5 mini by silence. Agreement among strong models fell from 31% of claims at turn one to 5% at turn four. With follow-up questions held fixed, it stayed at 28-33% through turn three and fell to 20% only at the closing turn. Stress runs showed the collapse guardrail escalating through its full intervention ladder when consensus was forced, but missing deference to one unusually strong seat. On idea-flow maps, the forced consensus appeared as parallel lanes of identical ideas, although quotation grounding and participation balance stayed at their maxima. We also reimplemented two close published systems and ran them on matched models, seats, and questions. A multi-persona debate cited only real records, yet none of its cited sentences carried a checkable quotation and it had no way to abstain. An author-memory fusion system ran only after its fact checker was split by seat to fit in context. Its fused answer kept about two-thirds of the fact-checked specialist claims, answered 36 of 60 question parts on which our seats had abstained, and showed the reader 19% of its agents' disagreements. In a computational pathology case study, six scientist-associated agents built distinct pipelines to predict MYCN amplification from whole-slide images. Across six base models, seat-by-model interaction explained 57% of internal variance, and internal rank poorly predicted external rank (r = 0.18 across 36 nominees). These results establish differentiated, source-traceable scientific reasoning as an operational capability and identify the base model as a parameter that must be reported with the protocol.