NTT reports preliminary results for AI-ready software traceability
The research links specifications, code and tests, with a small experiment distinguishing detected inconsistencies from wider production performance.

Edivro Company News · Original attributed summary. Source announcement: 2026-10-23.
NTT has described research intended to make existing software easier for AI systems to understand and modify. Its 23 October Japanese-language release explains methods for tracing relationships between specifications, design documents, code and tests, and for generating tests where coverage is incomplete.
What did the preliminary experiment find?
Using the ATRS sample application, researchers tested detection of 41 deliberately introduced documentation errors. NTT reports that its proposed method detected 33, about 80%, compared with 11, about 27%, using the stated retrieval-based comparison method.
A separate test found four of six deliberately introduced code bugs, along with three additional latent bugs. These are results from a defined preliminary experiment, not a general accuracy score for AI software engineering.
What comes next?
NTT plans evaluation on systems with different sizes and characteristics and intends to apply the techniques to actual development and maintenance projects. The work is scheduled for display at NTT R&D Forum in November.
The announcement establishes a research approach and initial findings. It does not show that the method has already replaced human review across production systems or achieved the same detection rates in every codebase.
Questions about this story
These answers summarise the article and refer to its original reporting period.
What did the preliminary experiment find?
A separate test found four of six deliberately introduced code bugs, along with three additional latent bugs. These are results from a defined preliminary experiment, not a general accuracy score for AI software engineering.
What comes next?
NTT plans evaluation on systems with different sizes and characteristics and intends to apply the techniques to actual development and maintenance projects. The work is scheduled for display at NTT R&D Forum in November.
Sources and links
Publisher
Edivro Company News


