TEE-Based Architecture for Auditable AI Grant Evaluation

A recent study introduces a Trusted Execution Environment (TEE) framework aimed at ensuring the auditability of AI-driven grant evaluations while safeguarding the model and scoring methodology. As government entities increasingly adopt large language models (LLMs) for decision-making assistance, a governance challenge arises: applicants should not be able to manipulate the model, while maintaining a process that is both contestable and accountable. This architecture employs remote attestation, enabling an external verifier to authenticate the model, rubric, prompt template, and input representation used, without disclosing model weights or proprietary scoring methods. Central to this framework is an attested evaluation bundle—an authenticated, timestamped document that connects the original submission hash, canonical input hash, model-and-rubric metrics, and evaluation results. The study also explores a scenario where the verifier can ascertain the correctness of the evaluation process.