Bachelor and Master Theses

Architecture documents and diagrams can become inconsistent with the software they describe. However, differences do not always indicate errors: a diagram may deliberately omit implementation details. This thesis investigates whether automated methods can detect meaningful contradictions while recognizing legitimate abstraction and incomplete evidence.

Possible Research Goal

Evaluate methods for identifying and locating contradictions between documented architecture, diagrams and implementation.

Possible Research Questions

1. How accurately can automated methods detect architectural contradictions across these artifacts?
2. Does combining AI with explicit mappings and deterministic checks improve detection?
3. Can these methods distinguish contradictions from consistent refactorings, intentional abstraction and insufficient evidence?

Possible Research Method

Students will conduct a comparative evaluation using public GitHub repositories:

1. Select projects containing versioned architectural documentation, diagrams and corresponding implementation.
2. Construct validated cases containing genuine inconsistencies, consistent changes and unresolved relationships.
3. Supplement these cases with clearly labelled, controlled changes that introduce contradictions or preserve consistency.
4. Compare direct AI assessment, conventional analysis tools and a combined approach.
5. Measure detection accuracy, false alarms, localization, supporting evidence and execution cost.

The initial target is 80 historical artifact bundles across 8–12 projects, supplemented with controlled variants and refined through a pilot.

Programming and debugging unfamiliar codebases.

Understanding dependencies, interfaces and modular structure.

Git, build tools and automated testing.

Writing scripts to extract, transform and compare software artifacts.

Knowledge of software architecture and modeling is required.

Experience with static analysis, diagram formats or LLM APIs.