The system evaluates claims across four core verdicts—correct, incorrect, gap, and unverified—by matching research findings against memory using three independent provenance flags (parametric recall, grounded research, and human review). When research confirms or contradicts an ungrounded memory claim, its verdict updates accordingly, while newly discovered facts absent from parametric recall are minted as gaps. If a human operator submits a manual ruling, the human-lock protection locks the verdict against automated model overwrites while still allowing future runs to attach supporting evidence.