AI COBOL-to-Java migration solves translation but not behavioral verification — here's the gap

Short summary

AI coding agents can now translate COBOL to Java quickly, but proving the migrated code preserves identical business behavior remains unsolved. Spot-testing misses subtle changes like fixed-point decimal vs floating-point rounding that compound across millions of transactions. The author distinguishes translation accuracy (largely solved) from migration completeness (capturing undocumented business rules and operational quirks) and proposes Business Rule Recall as a better metric than lines-of-code translated.

• •AI can translate COBOL to Java but cannot prove behavioral equivalence — the critical gap in legacy migration
• •Floating-point vs fixed-point arithmetic differences can cause silent financial discrepancies that surface as audit findings months later
• •Business Rule Recall — measuring behaviors preserved, not lines translated — is a more honest metric for migration completeness

Generated with AI, which can make mistakes.