Nivetha

AI can rewrite COBOL in minutes, but someone still has to decide whether to accept the Java. I worked on a COBOL modernization at a global finance company where every translated program was checked by hand. Diocle runs the original under GnuCOBOL and the Java on the same inputs and compares the output. Java that passes is attacked with inputs Muse Spark designs, and a repair agent fixes what differs. A model never certifies; the comparison does. Delivery is a pull request with an equivalence report, and CI re-runs the comparison on fresh inputs. On 15 synthetic programs, 4 of the 10 attackable translations that passed 1,000 generated records then differed under attack, on malformed numbers or the sign of a zero. In a blind test on COBOL from the public repository of Norway's retired national pension system, 2 of 4 testable programs certified, after repair of the Java, and 2 did not. This is testing on stated inputs under GnuCOBOL.