DON'T BE TRAPPED IN ILLUSIONS, Answering the AI Czar Project: Why LLM Regulation Cannot Succeed, and What Can

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The incoming AI czar faces a structural choice, not a political one. Large language modelscannot be functionally safety-certified per instance — this is a property of stochasticsystems, not a gap in current standards. Any regulatory framework written for them willtherefore be a process standard, will be received as a safety standard, and will fail where it matters. This paper argues that the productive path is not to certify the LLM but to certify thealternative: a class of deterministic, bounded-domain language models (DLMs) for whichper-instance certification is possible, and which bypass the data-center energy wall for their admissible task class. ST-T1024 is a published, Apache-2.0 reference standard for that class. The ask is not deregulation. The ask is redirection of certification effort towardsystems that can be certified at all.