Measuring Organized Answerability in Autonomous Robots Using Experience Architecture: A Reproducible Structural Evaluation Framework for RLBench

Conventional evaluation of autonomous robots emphasizes task success, return, completion time, and endpoint error. These measures determine whether a task was completed but do not reveal whether task-relevant discrimination, information access, goal orientation, multimodal integration, temporal continuity, and action consequence remained mutually organized during execution. This paper develops a robotics-specific operationalization of Experience Architecture, a typed constrained-dynamical framework for organized answerability. Each robot episode is represented by six normalized conditions--Difference, Availability, Orientation, Integration, Temporality, and Answerability--connected through a directed mutual-constraint graph. Their organization is summarized by the Experience Architecture Completeness Index, which combines enabling-condition adequacy, directed coherence, constraint satisfaction, and counterfactual answerability through zero-sensitive geometric aggregation. A reproducible RLBench protocol is defined for twelve manipulation tasks, one hundred demonstrations per task, five random seeds, three policy classes, and graded condition-targeted perturbations. The package implements matched intervention losses, task-specific goal diagnostics, grouped prediction, mixed-effects inference, bootstrap uncertainty, and machine-generated reporting. Six unit tests and an end-to-end synthetic pipeline verification on 8,100 artificial records confirmed executable metric computation, table and figure generation, provenance capture, and software integrity. The synthetic outputs are reported solely as implementation evidence: they are not RLBench observations and are not interpreted as robot-performance, construct-validity, or predictive-validity evidence. RLBench, PyRep, and CoppeliaSim were unavailable in the execution environment.