A Locksmith Loop paper reports 91.90 percent branch coverage on one internal COBOL program when validating agent-migrated code against the original with a deterministic parity oracle

Sources: paper, HN discussion
Summary: The preprint describes Locksmith Loop, a method that validates agent-migrated COBOL against the original program with a deterministic parity oracle. The 91.90 percent branch coverage figure describes one of three case studies, an internal production-like COBOL program, and the abstract reports nearly complete coverage for the two open-source programs. The corpus spans 430 to 4,114 source lines. The method instruments the COBOL source and the generated Java with mocks and runs both off-mainframe on commodity hardware.
Why it matters: The contribution is a deterministic oracle for agent output rather than a model score, so the check does not depend on a second model agreeing with the first.
Follow-up: Preprint with no peer review. Track whether the parity oracle is reproduced outside the authors' corpus.