Aelitium: Cryptographic Proof for AI Model Outputs

Aelitium offers a novel approach to ensuring the integrity and verifiability of AI model outputs. By creating tamper-evident evidence bundles, it allows independent verification of AI-driven decisions, addressing a growing concern about accountability and transparency in AI systems. The tool utilizes deterministic SHA-256 hashing, ensuring that the same AI output always produces the same hash, regardless of the machine used. This feature is particularly valuable for detecting model drift, where an AI model silently changes its behavior for the same prompt. Aelitium's ability to function offline further enhances its utility, eliminating the need for constant server connectivity. The capture adapter intercepts API calls, recording request and response hashes, and automatically writing the bundle. This automation simplifies the process of integrating Aelitium into existing AI workflows. The tool supports various AI platforms, including OpenAI and Anthropic, and offers features for streaming and signing. By providing a means to prove what an AI model actually said, Aelitium helps build trust in AI systems and reduces the risk of unintended consequences. The reproducibility check confirms the resulting hashes match, ensuring the integrity of the AI system. All tests also pass on two independent machines with identical hashes.