Spec2COBOLRot: An Agentic-AI Degradation Loop for Realistic COBOL Corpus Generation
COBOL remains widely deployed, yet representative corpora reflecting real production code are rarely available, limiting rigorous benchmarking of modernization approaches. We propose a systematic agentic AI pipeline for generating realistic COBOL programs, combining specification-driven generation with iterative degradation guided by patterns and complexity targets extracted from real production code. Here, realism is understood as structural fidelity to production code as captured by our metrics. We evaluate whether degradation reaches target complexity levels while preserving business behavior, and examine the limits of the approach, across three programs from distinct business domains....
| Authors | Jean-Baptiste Espinasse, Djamel Eddine Khelladi, Mathieu Acher |
|---|---|
| Primary category | cs.SE |
| Published | 21 September 2026 |
| Last updated at source | 21 September 2026 |
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.