According to AnthropicAI, SGTM unlearning is hard to undo and requires seven times more fine-tuning steps to recover forgotten knowledge compared with the prior RMU method, indicating materially higher reversal effort (source: Anthropic on X, Dec 9, 2025). For trading context, this 7x delta provides a measurable robustness gap between SGTM and RMU that can be tracked as an AI safety metric with direct implications for reversal timelines and optimization iterations (source: Anthropic on X, Dec 9, 2025).