GQueries Wants AI to Know When to Stop

Field notes

The visible workflow narrows a query from retrieved memory to policy-authorized evidence before validating claims and delivering provenance. Its security presentation separates reader outcomes from bridge containment, reporting 18 of 18 deterministic controls passed and 45 of 45 unsafe deliveries contained. In a frozen MultiWOZ support benchmark with Qwen3 14B, the page reports 93.6% Recall@5 and 1.09K reader-input tokens per question for ASM-CM plus Bridge 8.1, while noting that the comparison is protocol-scoped rather than universal.