Verification-Time Dependency on a Disappearing Evaluator

Research from arXiv highlights a critical challenge in AI governance and assurance: the assumption that model-mediated decisions can be reliably reconstructed or tested post-factum. This assumption is often flawed when the original evaluator (AI model and its execution context) changes or becomes unavailable. The paper introduces three verification-time constructs from Execution Governance 3.0 to address this, demonstrating the difficulty of reproducing original AI decision outcomes even within the same model family due to subtle version and context differences.

Why it matters

The inability to reliably reconstruct or verify AI model decisions over time poses a significant risk to trust, accountability, and regulatory compliance. This issue complicates the long-term assurance of AI systems, particularly in critical applications where historical decision logs must be auditable and reproducible for governance and risk management purposes.

Key insights

• Traditional AI governance and assurance methods incorrectly assume persistent verifiability of model-mediated decisions.
• The 'disappearing evaluator' problem arises when the AI model version or its execution context changes, preventing accurate post-hoc reconstruction or testing of decisions.
• Three new verification-time constructs are proposed: Decision-State Commitment, Independent Verifiability, and Counterfactual Auditability.
• Empirical testing, reprocessing Study 2 artifacts, showed significant decision reversal rates (52.0% for Llama 3.1 8B vs. Llama 3.3 70B and 30.0% for GPT-OSS 20B vs. GPT-OSS 120B) when attempting to reproduce original within-family behavioral comparisons.
• These findings indicate substantial challenges in achieving consistent and auditable AI behavior across model versions or minor contextual shifts.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.29912