ARTICLE B --- The Alchemy of LLMs: A Field Without a Foundation, and One Attempt to Build One

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 current paradigm of large language model research is a random walk through acombinatorial space too large to reason about directly. Progress is real but unsystematic: results are found by luck, swarmed by hundreds of follow-up papers, and abandoned whenthe next result appears. The field has engineering without science — no first principles, no predictive interpretability, no formal guarantees — and it is now hitting physical and conceptual walls simultaneously. This paper argues that the paradigm is not a foundation but a search heuristic, and that its continuation is guaranteed by funding incentives rather than by evidence of convergence. It then presents a worked counterexample: ST-T1024, a deterministic, bounded, sovereign architecture for embedded inference, with a formal specification, a reference emulator, and an agentic extension. The counterexample is not a benchmark victory. It is an existence proof that a different direction is already running.