Ceteris Paribus and Sequential AI Inference: A Framework for Evaluating Precision and Runtime Governance
Year: 2026
Type: preprint
Abstract: A high sequential artificial intelligence (AI) benchmark score measures success under aparticular grading rule, not necessarily valid workflow resolution, accurate outcomes, or authorizedconsequential execution. This position and methodology paper distinguishes five operationalmeasures—task completion, valid workflow resolution, adjudicated accuracy, claim precision,and state precision—from a separate governance-assurance dossier. Motivated by conditionalevidentiary warrant (ceteris paribus), we specify three noninterchangeable governance decisions:independent Governance Compilation Model (GCM) validation establishes proposal eligibility; theAI Governance Control Plane (AGCP) acquires authoritative authorization factors and mediatescommitment; and AGCP independently verifies target state and evidentiary closure. We statetrust and correspondence obligations, define boundary-specific decision precision, recall, coverage,and cost, and provide a complete synthetic transaction and a staged experimental protocol.The protocol includes a narrow feasibility pilot and a comparison of LLM-declared obligationswith trusted obligation attachment. MedAgentBench v2 results motivate the study but do notestablish its proposed governance properties. No new intervention results, implementation-levelproof, or clinical assurance are claimed.
Author: Willis, John
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225746938
Cites:
Cited by:
Related to: