When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows

Antibody lead optimization calls for a small, bounded set of edits to an existing candidate: substitutions, but also insertions and deletions. Edit-based generative models are the only ones that allocate such an edit budget without fixing the edit positions, the edit count, or the output length in advance. However, the existing approaches Edit Flows and EvoFlows did not release code or complete training specifications. Here, we show that both methods follow the same underlying process — edits firing one at a time, at learned rates, in continuous time — the pure-jump case of generator matching over finite sequences. With \textit{EditJumps} we introduce the first open implementation of this framework, with a single generalist antibody editor trained on 1.66M Observed Antibody Space homolog pairs to propose homolog-like variants of a seed sequence, editing unseen leads zero-shot, without the per-family retraining original approaches require. At a comparable edit budget (approximately 5 edits per sequence) EditJumps generates more diverse variants than evotuning and EvoDiff-MSA while matching them on distributional fidelity. Reconciling the published and replicated edit statistics was only possible after we identified an undocumented hyperparameter controlling how many edits a sequence receives, and we find most of the published evaluation metrics to be underspecified, some sensitive enough to reference-set size to reorder methods. The code, tests and run configurations are published at: https://anonymous.4open.science/r/editjumps-BA56/.