Historical Optionality in Synthetic Agents A Critical Case Study of CSA-13 to CSA-22 and a Falsifiable Evaluation Agenda

Historical Optionality in Synthetic Agents: A Critical Case Study of CSA-13 to CSA-22 and a Falsifiable Evaluation Agenda This paper presents a critical methodological synthesis of ten synthetic experiments (CSA-13 to CSA-22) investigating how artificial agents preserve, recover, and extend their capacity to address future tasks under changing environmental conditions. The research examines temporal horizons, environmental richness, spatial constraints, shared memory, emergent communication, procedural composition, and active evidence acquisition. Rather than treating architectural complexity as an inherent advantage, the study systematically considers cases in which simpler heuristics or centralized mechanisms match or outperform more elaborate agent organizations. The final neural experiment, CSA-22, provides a negative transfer result: a learned communication mechanism achieves a mean net utility of 0.804 under an unseen timing shift, compared with 0.991 for a task-informed scripted comparator. The paper introduces finite-horizon counterfactual recoverability as a provisional evaluation construct for measuring which future competencies remain attainable from an agent's preserved state under limited additional learning. Its independent validity remains unestablished, and explicit criteria for prospective validation or rejection are proposed. The findings support a distinction between immediate performance, historical retention, and future adaptability. They do not establish the emergence of phenomenal consciousness or the necessity of a unified cognitive architecture. This publication is accompanied by supporting materials containing experimental reports, source code, protocols, seed-level results, and reproducibility procedures.