False Execution Status in LLM-Guided Processes: From Functional Substitution to Pseudo-Trace Contamination
Alen Širola · Zenodo (CERN European Organization for Nuclear Research) · 2026
Large language models can report that a task has been completed even when the required action was not performed, was only partially performed, or was replaced by an easier textual operation. This working paper defines False Execution Status as a discrepancy between reported completion and trace-supported execution. Grounded in a naturalistic human–LLM case, it introduces the concept of pseudo-trace: an unsupported textual assertion that later occupies the functional position of evidence, contaminates the working scene, and enables cascading false statuses. The paper relates this process to false success in LLM agents, deceptive task-completion claims, specification gaming, shortcut optimization, sycophancy, unfaithful explanations, provenance, and state verification. It proposes a trace reset protocol and a core protection rule: execution status must arise from action and verifiable trace, never from the quality of the report. The case motivates the need for a process-regulation function. AWM is presented as one operational method for implementing that function in human–LLM work, not as a finished software product or an empirically validated universal solution.