Abstract
This thesis compares the reliability and practical applicability of different techniquesfor verifying the behavioral equivalence of two pure functions.
Five approaches are evaluated on a common dataset in Rust: exhaustive testing,
property-based testing with proptest, bounded model checking with Kani, bounded
model checking with loop contracts, and deductive verification with Verus. Each
function group under test consists of an original implementation, a modified variant
with identical behavior, and a deliberately altered variant with divergent — that is,
faulty — behavior. Every approach was expected to establish equivalence between the
original and the correct variant, and, when comparing the original with the incorrect
variant, to at least fail, or better still, to produce a counterexample.
The dataset comprises 103 function groups in four categories that differ in the size of
the input space and in their use of arrays, strings, and library functions. For deductive
verification, the proofs are not written by hand but generated with Claude Code and
Fable 5 and checked by Verus. The techniques are assessed by correctness (no false
equality), completeness (no false inequality and), applicability to the four categories, and
effort. Each result is classified as proven, empirically shown, or assumed. From this, the
thesis derives which technique suits which kind of function and which level of assurance.
Keywords: equivalence checking, formal verification, bounded model checking, property
based testing, deductive verification, Rust, Kani, Verus, proptest, pure functions
| Date of Award | 2026 |
|---|---|
| Original language | German (Austria) |
| Awarding Institution |
|
| Supervisor | Erik Pitzer (Supervisor) |
Studyprogram
- Software Engineering
Cite this
- Standard