Research
Differential Fault Injection Tests Whether LLM-Modernized Code Preserves the Original's Failure Behavior, Not Just Its Happy Path
Validation of LLM-driven Fortran modernization typically checks nominal executions, leaving open whether the rewrite preserves how the original responded to faults, perturbations, and reduced precision. This harness instruments the shared self-consistent-field driver of GAMESS at twelve sites and applies identical deterministic faults to both implementations across more than 2,200 runs. Original and modernized kernels agreed in all 200 paired injections and a synchronization change composed cleanly in all 40 pairs — but the campaigns also exposed phase-dependent parallel deadlocks and false convergence under reduced precision that nominal testing would never have surfaced.
↳ Follow the thread