📄 中文摘要
该研究提出了一种使用非对称密钥签名（EdDSA）对每个交互阶段进行正式签名的方法，以增强 LLM 提示实验的可验证性。当前在提示工程和上下文工程的讨论中，证据链的可靠性较弱，难以实现 100% 的验证。项目的灵感源于 Michael Reeves 的一段 YouTube 短视频，其中他通过直接编辑过去的对话文本来“误导” LLM，导致其逻辑崩溃。该方法旨在解决这一问题，确保每个交互阶段的完整性和可追溯性，从而提高 LLM 的探索和应用的可信度。
严格验证 LLM 提示实验的加密方法提案
📄 English Summary
Proposal For Cryptographic Method to Rigorously Verify LLM Prompt Experiments
This research proposes a method for formally signing each stage of interaction using asymmetric key signing (EdDSA) to enhance the verifiability of LLM prompt experiments. Current discussions on prompt engineering and context engineering reveal a weak chain of evidence, making 100% verification challenging. The project was inspired by a YouTube short from Michael Reeves, where he 'gaslights' the LLM by directly editing past dialogue, causing a logical breakdown. This method aims to address this issue by ensuring the integrity and traceability of each interaction stage, thereby increasing the credibility of LLM exploration and applications.