Title: Calibrated Deference in Closed-Loop Molecular Design

We introduce Cognitive Loop via In-Situ Optimization (CLIO), an intelligent agent that integrates a recursively updated belief-state graph with a continuous plan-then-act cycle. This architecture enables a form of reasoning we define as calibrated deference: the ability to detect when existing tools or underlying assumptions are insufficient, adjust strategies accordingly, and formulate mechanistic hypotheses to direct experimental adjustments.

To validate this approach, we engaged CLIO in a closed-loop human-AI collaboration aimed at designing a negolyte for aqueous organic redox flow batteries (AORFBs). In this partnership, CLIO spearheaded the proposal and interpretation phases, working closely with chemists who handled synthesis, characterization, and strategic decision-making.

Over three iterative rounds involving 17 distinct candidates, CLIO identified a leading phosphonate-based compound. Subsequent characterization demonstrated a 130 mV enhancement in redox potential compared to established literature baselines. However, further analysis uncovered unexpectedly low electrochemical reversibility—a performance regression that automated property predictors had failed to anticipate.

In response, CLIO developed alternative mechanistic explanations, identified key diagnostic tests to distinguish between them, and pinpointed phosphonate-potassium ion pairing as the root cause of the failure. It then recommended substituting the phosphonate group with a sulfonate. The final compound exhibited significantly better electrochemical reversibility while retaining a 90 mV advantage in redox potential, thereby successfully completing the design-make-test-redesign cycle.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC