Governing Two Primary LLM Review Streams and a Separate GPT-5 Audit in Evidence Synthesis: Source-Based Human Reconciliation in a Bangla Visual Speech Recognition Case Study

Large language models (LLMs) are increasingly used in evidence synthesis, but model agreement cannot substitute for source verification. We retrospectively examine a human-governed workflow embedded in a Bangla visual speech recognition evidence map. Two primary LLM streams assessed the same paired records in separate sessions using retained workbook instructions. The paired pool contained 17 records: 16 Bangla/Bengali records and one Assamese record (BA-017). After restriction to Bangla/Bengali by language scope, 16 paired records remained; all nine included Bangla reports were retrieved in full. Exact title/abstract decisions matched for 16/17 records in the full pool (94.1%; kappa = 0.887) and 15/16 in the Bangla set (93.8%; kappa = 0.881). Operational retrieve decisions matched for all records in both sets, and both streams identified the same nine included and seven excluded Bangla records. Separately, GPT-5 recoded 81 post hoc source-verifiable audit cells; 76 matched the reconciled provisional ledger (93.8%). Humans source-checked all 81 cells, and the five discordant provisional values changed after source adjudication. Because adjudication was not blinded to discrepancy status, the audit comparison is not an accuracy or independent-validation estimate. This bounded case contributes a role-and-record framework for auditable LLM-assisted evidence synthesis while retaining human authority over final source-based decisions.