Multi-agent deliberation systems that operate across multiple sequential rounds face a structural problem with no analog in single-pass AI systems: evidence introduced late in the deliberation sequence cannot receive the same adversarial scrutiny as evidence introduced at the start. A Contrarian agent that receives a new evidence node in Round 3 has no opportunity to challenge it; a Methodologist that has not assessed a node cannot bound it; a Synthesizer that receives an unassessed node has no confidence ceiling to enforce. We describe the RoundAware Evidence Admission Protocol (REAP)—a tiered evidence governance framework integrated into the Augle seven-agent deliberation ensemble that applies round-specific admission rules to evidence nodes introduced at different stages of deliberation. Round 1 is unrestricted. Round 2 requires a materiality justification tag; the Methodologist must assess any late-introduced node before it may enter the evidence registry. Round 3 evidence is inadmissible to the current session and is routed to a structured future-session gap record. The protocol is informed by the evidentiary rules of adversarial legal proceedings, in which the timing of evidence introduction is a first-class procedural concern, not an afterthought. We describe the protocol design, its integration with the confidence propagation framework and evidence node registry, and the automatic fallback rules that govern unassessed late-evidence nodes. Companion papers: Kelly, C. Zenodo DOI 10.5281/zenodo.20619123. Kelly, C. & Saxena, S. (2026). “Source Verification as a First-Class Architectural Layer in Multi-Agent AI Deliberation Systems.” Zenodo, June 2026.
Kelly et al. (Mon,) studied this question.