Harness Meets Diffusion: Single-ReAct-Guided Crystal Generation and Knowledge-Grounded Superconductivity Search

Scientific crystal discovery requires generators that can adapt their search to explicit property targets. We introduce a single-ReAct controller for a frozen LLM-diffusion crystal generator. The controller uses batch summaries and retrieved materials knowledge to guide subsequent LLM proposals. Selected candidates then undergo diffusion refinement. In open-ended generation, it raises the stable, unique, and novel fraction ($\mathrm{SUN}$) from $0.2131\pm 0.0015$ to $0.3436\pm 0.0749$ across three matched seeds, with a gain in every seed and a decline in reference-distribution fidelity. We then apply the controller to superconductivity screening using a frozen $T_c$ predictor and a separate structural predictor for evaluation. In comparisons with the same generator and candidate budget, the controller achieves $\mathrm{SUN}=0.455\pm 0.040$ across four seeds, compared with $0.257\pm 0.007$ for best-of-$k$ and $0.241\pm 0.009$ for verifier-only selection. Across six seeds, it generates $2.26\times$ as many candidates with predicted $T_c\ge 20$ K and nearest-training-composition Jaccard distance at least $0.20$ as a knowledge-only policy. Adding a superconductivity- tuned generator further increases the joint stable, unique, novel, and predicted-$T_c\ge 20$ fraction from $0.1723$ to $0.2843$ in a matched run. These results show that feedback-driven control improves the computational screening yield of an existing crystal generator across two scientific objectives.