PLaND Path to Least Non Determinism

Abstract

Language-model workflows can spend tokens repeatedly applying decisions that explicit code can perform. We present Path to Least Non Determinism (PLaND), a methodology for revising an English standard operating procedure (SOP) by replacing selected model decisions with code while retaining model fallback. Two reusable skills specify baseline construction and development-guided revision; a fixed comparison contract preserves the model, evaluation conditions and complete English fallback. We evaluate the resulting classification packages on LEDGAR legal clauses, CFPB consumer complaints and SpamAssassin email using Gemini 3.5 Flash Lite. Each dataset contains 500 development, 1,000 selection and 500 reserved final-test cases. Three paired executions reuse the same cases at each reached stage. Acceptance requires an 80% accuracy floor for both the baseline and hybrid, a paired 95% accuracy-difference interval with lower endpoint at least -2 percentage points, at least 5% token reduction, and a strictly positive lower token-reduction endpoint. LEDGAR passed all final-test repeats: mean accuracy changed from 94.93% to 95.80%, with 74.56% mean token reduction. SpamAssassin also passed: mean accuracy changed from 97.67% to 97.20%, with 28.13% mean token reduction. CFPB was rejected at selection because every hybrid repeat missed the accuracy floor; mean accuracy changed from 79.73% to 79.27%. Its final test was not evaluated. These results support selective code execution under the stated tolerances. They do not establish comparative superiority, reliable autonomous rule discovery, or production-workflow performance.