Wrapping an LLM in a five-stage deterministic control loop took constraint satisfaction from ~26% to ~96%

arXiv 2609.19710·medium signalXBlueskyLinkedInCopy link

For any feature that must hit a checkable numeric target (word count, readability band), the paper compares single-shot prompting against a generate-evaluate-adjust-archive-analyze loop where the model only writes and edits while deterministic code makes every accept decision and rejects any edit that drops source entities, numbers or keywords. Over 114 single-shot jobs and 240 closed-loop runs on four commercial models, single-shot met the target 21.1-31.6% of the time and the closed loop 92.5-98.8%, averaging under two edit rounds at 0.92-0.93 recall-based fidelity. The transferable rule the authors state plainly: declare the acceptance condition as code, bound the model to local edits, gate every edit on a content check.