When the Next Step Is Not One Step: Distribution-Aware Execution Modeling for Concurrent Go Programs

In three lines7B model fine-tuned to predict next step in concurrent Go programs by learning event distributions rather than single labels. On 798 predictions from real bugs (CockroachDB, Kubernetes, gRPC, etcd), achieves 36.2% accuracy with <1000 traces, outperforming Gemini 3.5 Flash zero-shot (34.8%). Dataset, adapters, and tooling released.

Summary generated by Claude — human-verified