From source code to a proven system.

Point LegacyPhoenix at the source tree. Parsers and scanners read programs, transaction definitions, screens, dictionary and database definitions, command procedures and build files, and write one model. Files with no parser are counted and named, so you see on the first day what is covered and what is not.

Open the navigator. The system is shown as a network of programs, data, screens, transactions and operating system objects. Start from any of them, follow any kind of link in either direction, search by name, and open the source line behind any fact.

Generators write everything that follows from the model by rule: data classes, the database schema, transaction flow, screen templates and the application's wiring. Units that have no generator are translated by AI, one at a time, from the original source and the slice of the model around it. A unit with no translation is reported as untranslated. It is never passed off as finished.

Scenarios recorded from the original application are replayed through the new one, and outputs, status codes, messages and data are compared exactly. The report states how many scenarios match and how much of the system they exercise.

Each difference becomes a work item tied to the parts of the model that can have caused it. The generator, the runtime or the translated unit is fixed, and everything is run again from clean state. The loop ends when no difference remains and the agreed coverage is reached. You accept the proof. That is the third decision.