The paper compiles machine-readable standard operating procedures into executable pseudocode and uses a program-guided stack-machine runtime to page the active frame while an LLM performs semantic execution. In a three-arm SOPBench study covering six models, compiled text did not significantly hurt performance and improved results by up to 16.0 points where official prose underperformed. Runtime guidance helped strong models but harmed weak ones. On the Bank task, the three primary arms increased from 70.4 to 86.4 and 92.8, with 100% refusal correctness. The authors recommend compiling SOPs first and enabling active-frame paging only after checking model-level state discipline.
No heat snapshots are available in the last 24 hours.