Intel / article

Spec2COBOLRot: An Agentic-AI Degradation Loop for Realistic COBOL Corpus Generation

Source: arxiv.org
Author: Jean-Baptiste Espinasse (DiverSe), Djamel Eddine Khelladi (DiverSe, CNRS, IRISA, KHORA), Mathieu Acher (INSA Rennes, IRISA, DiverSe)
Date: Published Sep 24, 2026

Context

Generating complicated COBOL code does not automatically create a realistic test of legacy-software modernization. Spec2COBOLRot starts with programs generated from specifications, then repeatedly alters them toward structural patterns and complexity targets drawn from production code. The aim is to supply test material where representative real programs are scarce.

Across three programs, the authors report valid syntax and movement toward the targeted complexity. But the process did not always preserve business behavior, and matching structural measurements could produce code with an implausible maintenance history. The finding separates two requirements for useful synthetic test data: resembling old code's structure and faithfully representing the business operations that modernization tools must preserve.

Key takeaways · AI-distilled

The pipeline combines specification-driven generation with iterative degradation guided by patterns and complexity targets extracted from real production COBOL.
Across three programs from distinct business domains, the pipeline reliably produced syntactically valid code and moved it toward realistic structural complexity, with realism defined by the authors’ own structural metrics.
Targeting structural metrics independently of business logic risks programs whose complexity does not reflect a plausible maintenance history; the authors propose simulating a development history from scratch instead.

Read the source arxiv.org