Synthesizability is per-molecule; cost is per-library: Hub-Batching for Reaction-GFlowNets

Synthesizability-constrained molecular generation is evaluated one molecule at a time, but a laboratory makes a library. The cost of a library is not additive: compounds sharing an intermediate are cheaper together than apart, because the expensive part of making them has already been paid for. We therefore measure distinct high-reward molecules delivered per fixed reaction budget. Under this readout, a trained reaction-GFlowNet already contains the information needed to design a cost-efficient library: state flow at pre-terminal intermediates, recovered from a frozen checkpoint with no retraining or reward modification, reveals which cores are worth diversifying around. This procedure, which we call hub-batching, mirrors scaffold decoration in hit optimisation: make analogues around a high- value core so the cost of building it is amortized across many structurally distinct compounds. Hub-batching delivers a median 2.48× more distinct high-reward molecules than reward-ranked selection across three reaction- grounded generators and four targets (23 of 24 comparable conditions; no case in which the baseline wins or ties), and, where only the selection rule differs on an identical candidate pool, 2.51× against a cost-optimal solver. That solver is cheaper per candidate purchased, and one scoring oracle failed calibration against property-matched decoys. Reaction-grounded generation earns its keep at the library level even where per-molecule synthesizability does not distinguish it.