Beyond Task Success: CTVI, a Multi-Axis Methodology for Agent-Trajectory Validity and Evidence Admissibility, with a Reproducible Semi-Synthetic Evaluation

Abstract

This repository contains the preprint of the CTVI framework (Commit-Trajectory Validity Index), a methodology for evaluating agent trajectories beyond task success by integrating outcome validity, evidence admissibility, and reproducible computational auditing. The paper introduces a multi-axis measurement architecture, an operational record schema, and a semi-synthetic evaluation based on AgentDojo-derived trajectories. The accompanying materials document the methodology, experimental design, reproducibility procedures, and study limitations. This record is released to establish a permanent scholarly archive and a citable DOI while the manuscript proceeds through formal peer review.