Certifying Repeatable Entry-Replay Witnesses for COBOL-to-Java Migration
Source
arxiv.org
Author
Andras L. Ferenczi, Mariya Bessonov, Allwyn Ponnudurai, Chandrakanth R. Mamillapalli, Krishna K. Maddula, Krishna K. Lingamneni
Date
Why it matters
Describes a way to generate behavior-equivalence tests for agentic legacy migrations by replaying recorded external operations against a surrogate program. It addresses the shortage of comparison test data.
Key takeaways · AI-distilled
To generate comparison tests for COBOL-to-Java migration, the authors run a surrogate COBOL program that mocks external systems with a chronological tape of recorded operations, then certify test cases by observing its execution trace.
On a 33,429-line COBOL program they produced 849 synthetic end-to-end tests covering 78.1% of source branches and 95.7% of main paragraphs; the generated Java matched the surrogate on all 849, and the campaign surfaced 61 generator defects.
When 450 defects were seeded into the Java, the gate caught 47.6%: every deleted write and deleted paragraph call in executed code, but none of the rounding and precision changes, a blind spot for validating AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →-written migrations.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.