Deterministic Code Deciding Every Accept Took Numeric Constraint Satisfaction From 21% to 98%

arXiv 2609.19710·medium signalXBlueskyLinkedInCopy link

arXiv 2609.19710 (17 Sep 2026) addresses software features that embed an LLM but must hit a checkable numeric target such as a word count or readability grade band. Its five-stage loop (generate, evaluate, adjust, archive, analyze) calls the model only to write and edit, while deterministic code compares a composite readability value against the target band, rejects any edit that drops source entities, numbers or keywords, and makes every accept decision. Across 114 single-shot jobs and 240 closed-loop runs on four commercial models, single-shot prompting hit the target 21.1% to 31.6% of the time and the closed loop 92.5% to 98.8%, within two edit rounds on average at recall-based fidelity of 0.92 to 0.93.