HiRes: Stress-Testing Inspectable Precedent Memory for Reaction Condition Recommendation

Reaction-condition models are components of synthesis-planning agents, but their reported accuracy can depend more on missing labels and permissive random splits than on chemical generalization. We present HiRes, a reaction representation that supports learned condition heads and retrieval of inspectable training precedents, together with a stress test prompted by peer review. We retrain five systems on the released USPTO-Condition split and four harder scaffold, reaction-class, patent- family, and temporal splits, evaluating each with and without the absent-condition class under one locked-test harness. HiRes-Hybrid leads the released split under both protocols and improves on a pure learned-retrieval variant in all ten split– protocol cells, but REACON and RCR lead most harder splits and a matched DRFP k-NN remains a strong baseline. A frozen six-arm ablation attributes measurable gains to reaction-center and GNN structural streams, not atom-map bias or an engineered/DRFP stream. Finally, a blinded 30-query study finds 50 of 60 retrieved precedents useful, while finding no usefulness advantage over DRFP. The result is both a competitive precedent-based recommender and a clearer account of what current condition benchmarks can and cannot establish.