Checking Actions to Audit Explicit Price Agreements

An AI monitor must distinguish an agreement that was implemented from a proposal, a canceled deal, or coincidentally matching actions. I evaluate whether extracting evidence and checking it with deterministic code improves this distinction in scripted pricing records involving two sellers. Two studies compare direct model judgment, agreement extraction with trusted action logs, and extraction of both agreements and actions followed by verification. Study 1 reuses 16 wording families; Study 2 introduces 12 new families with several changes to an agreement. Across 28,224 model calls, I test Qwen2.5 7B and 32B and Mistral 7B on JSON, table and prose records under clean, misleading and neutral notes. On the main comparison on prose records without misleading notes, joint extraction improves balanced accuracy from 68.4% to 78.1% for Qwen 7B, 99.0% to 100% for Qwen 32B in Study 1, and 65.0% to 88.0% for Mistral 7B and 93.8% to 100% for Qwen 32B in Study 2. However, smaller models often violate the required output schema, especially with JSON inputs. All 1,640 invalid outputs remain failures. The results support evidence extraction plus verification for this narrow task, while exposing a reliability problem that depends on the input and output format. They do not establish detection of spontaneous collusion or robustness to compromised monitors.