Scarf (Self-Contained Application Refactoring) is a benchmark suite for evaluating AI agents' ability to migrate enterprise Java applications across Jakarta EE, Quarkus, and Spring while preserving functionality, idiomatic patterns, and architectural integrity.
Applications
Layers
Frameworks
Tests
102
6
3
1,331
All applications have been manually converted and verified by experienced developers.
Consolidated multi-module coffee-shop (Jakarta, Spring) into single Maven modules for strict architectural parity across frameworks; Dockerfile updates for PetClinic
Domain-Driven Design cargo shipping tracker with Jakarta Faces, CDI, JPA, REST, Batch, and JMS
Coffee Shop
Event-driven microservices (Orders, Barista, Kitchen) via Apache Kafka with MicroProfile
DayTrader
High-performance stock trading benchmark with JPA optimistic locking and transaction management
PetClinic
Veterinary clinic management with Jakarta Faces (PrimeFaces) and complex JPA relationships
RealWorld
Medium.com clone (Conduit) with MicroProfile JWT, JAX-RS, and Testcontainers integration tests
Citation
@misc{scarfbench,
author = {Krishna, Rahul and McGinn, Bridget and Pavuluri, Raju},
title = {{ScarfBench}: A Benchmark for AI-Driven Enterprise Java Framework Migration},
year = {2026},
howpublished = {\url{https://scarfbench.info}},
}