Towards Verifiable AI with Lightweight Cryptographic Proofs of Inference

This project addresses the critical 'Verifiable Inference' problem—the ability to prove a cloud provider actually ran the model they claimed they did. While ZKML (Zero-Knowledge Machine Learning) projects like Modulus Labs, EZKL, and RISC Zero focus on full ZK-SNARKs (which are historically slow for LLMs), this project targets a 'lightweight' middle ground. The defensibility is currently low (4) because, despite 7 forks suggesting academic interest, the project has zero stars and represents a research-stage implementation rather than a hardened tool. The moat is purely algorithmic; if the underlying protocol is effective, it will likely be absorbed into larger frameworks or standardized. The primary threat comes from two sides: 1) Hardware-based verification (TEEs/Confidential Computing like NVIDIA's H100/H200) which provides similar guarantees with less software overhead, and 2) Frontier labs (OpenAI/Google) implementing proprietary watermarking or lightweight commitment schemes that satisfy 90% of user needs. Market consolidation is expected as verifiable inference is a 'feature' that will eventually be integrated into model serving stacks (like vLLM or Hugging Face TGI).