SIGNAL
Enterprise AI's auditability is set by what the model is not allowed to invent, not by how smart it is. The system that just won the cleanest head-to-head of the year does not let the language model pick its own query path, run its own tools, or improvise its own math. A fixed, pre-approved program does the work, and a small model only interprets the question. The result is replayable in a way free-range agents' answers are not.
That pattern shows up in a 2609.03209 paper. Three small models given free rein over SQL and tool calls scored zero across 330 runs on the same test set, while a configuration that lets the model interpret the question and hands the analytics to a pre-approved program matched 110 of 110. The winning setup moved the model's creativity from the data path to the question.
The stakes land in any company running an LLM on operational data: if the AI invents its own query path, can the auditor see what it did? The authors flag their own result as configuration-specific, not a verdict on the field. The design lesson is portable. Auditability is set by where the model lives in the stack and what it gets to decide. Vendor pitches that lead with model size and benchmark scores leave the control surface off the page. That is the page that matters.
Mechanism: the boring configuration won because the interesting one could not prove what it ran. Auditability travels with the policy, not the planner.