Reusable Neural Guidance for Petri Net Alignments with Replay Verification

Alignments explain deviations between an event trace and a process model, but difficult traces can make exact computation slow. This paper studies whether one pretrained neural scorer can be reused to propose alignments for different Petri nets and activity vocabularies. LARA (Learned Alignment with Replay Assurance) encodes the Petri net and the complete trace, scores possible moves, and constructs a candidate using a decoder that tracks the marking. Independent replay checks executability and computes the candidate cost. Optional exact search establishes optimality or supplies a replacement. We explain the neural machinery, synthetic training procedure, and verification contract. Across three training seeds, greedy decoding produces 90.8% optimal candidates on the 512-row synthetic test set (sample standard deviation: 0.3 percentage points), which can be compared with 86.9% without guidance. All candidates pass replay. An optional single-log repair procedure raises the reference checkpoint’s rate from 91.0% to 97.3%. The largest learned benefit concerns duplicate activity labels. Tests exclude training-isomorphic graphs, include cyclic structures, and evaluate held-out cases from three public logs. They reveal inconsistent structural gains and no real-log quality advantage from learned rankings. Common-boundary timings compare per-trace methods, while established approximations remain stronger than the greedy candidate policy. Certified mode runs exact search on every trace. The results support a limited role for reusable scoring, with guarantees supplied by process semantics.