I scanned 30 popular AI projects for tamper-evident LLM evidence. 0 had it (github.com)

A recent assessment of 30 major AI projects uncovered a critical gap in tamper-evidence systems for large language models (LLMs), revealing that none included mechanisms for creating independently verifiable execution evidence. The tool "Assay" offers a solution by generating cryptographically signed receipts that third parties can verify offline, thereby enhancing accountability in AI systems. This innovation comes as regulatory pressures increase, particularly with the upcoming EU AI Act, which mandates automatic logging for high-risk AI systems. Assay differentiates itself by providing not just logging, but an integrity-checking system that establishes if evidence has been tampered with while also verifying compliance with governance checks. Its implementation is straightforward, requiring only a few lines of code to produce a detailed evidence pack that enhances trust in AI operations. The tool’s ability to expose honest failures and maintain transparency in audits is crucial for building robust AI governance frameworks, especially as the demand for accountability in AI applications grows in line with regulatory requirements.