NTT reported on 23 October a method to help AI work with existing software. It groups specifications, designs and code by the functions they describe, then reconstructs downstream information from earlier development stages. Bringing descriptions to a comparable level of detail helps it map dependencies and check consistency.[1]
Seeded design inconsistencies detected
| Method | Detected out of 41 |
|---|---|
| NTT method | 33 |
| Ordinary retrieval-augmented generation | 11 |
The company also describes generating tests from missing combinations of functions and conditions. It plans further evaluation across real projects and programming languages. The experiment used ATRS.[1]
ATRS supplies the test setting
ATRS stands for Airline Ticket Reservation System. Its repository identifies it as a reference application for learning the Macchinetta development framework and libraries. That makes the setting concrete: a sample software system whose related assets can be examined together.[2]
Following a change in both directions
The underlying engineering need is older than generative AI. NASA’s traceability guidance describes links from requirements to design and back again, so teams can check that requested behaviour is implemented and that design elements have a reason to exist.[3]
Those links also help a team work out what a change affects. Consider an illustrative cancellation rule: its written requirement, implementation and tests need to stay aligned. A useful map lets an engineer start at the rule or at the code and find the corresponding material.[3]
Sources & further reading
- NTT Transforms Existing Software Assets into AI-Ready Development Assets NTT · 23 Oct 2026 · Accessed 29 Oct 2026
- Airline Ticket Reservation System (ATRS) Macchinetta · Accessed 29 Oct 2026
- SWE-059 - Bidirectional Traceability Between Software Requirements and Software Design NASA Software Engineering Handbook · Accessed 29 Oct 2026