A Hard Verbatim-Evidence Gate as the Precondition for Multi-Agent Claim Verification: Design, Deployment, and Empirical Boundary Conditions

Multi-agent large language model (LLM) pipelines verify claims by having model instances critique one another, but agreement reached through opinion carries no guarantee that any evidence supports it, and prompt-enforced grounding reduces unsupported assertion without preventing it. This study evaluates a structural alternative: a verdict may be Supported or Contradicted only if it quotes a span that pipeline code finds verbatim in the retrieved evidence; otherwise it is rewritten to Unverifiable. The gate was built into Aletheia, a deployed LangGraph verification service, and compared with a single-LLM baseline and an otherwise identical ungrounded ablation on SciFact and FEVER, using one frozen corpus, seeded stratified samples of 100 claims, and paired significance tests. On the deployed model the grounded arm ties the baseline on accuracy (79.0 versus 79.0 percent) and leads on catch rate (96.6 versus 93.1 percent) and false agreement (6.1 versus 10.5 percent), but neither interval excludes zero. On an 8-billion-parameter model the catch-rate gain is significant: 82.8 versus 60.3 percent, +22.4 points, 95 percent confidence interval [+12.1, +33.3]. On FEVER's paraphrased claims the ungrounded ablation is significantly more accurate (85.0 versus 77.0 percent). Re-checked from traces, 51 of 51 asserted FEVER verdicts are verbatim-backed, yet five are wrong: the gate guarantees a faithful quotation, not a correct inference. Structural grounding buys accuracy for weak models and auditability for strong ones — a boundary condition, not a leaderboard win.