Pilot results

Phase 4 applied SpringGraph to four applications. The full write-ups are in the Evaluation section of this site.

Application	Shape	Validated	Boundaries declared	False positives found
examples/pilot-boot	Boot 3.5, JPA, 7 files	2 of 6	2 (Clock, ObjectMapper)	1: @Value parameters reported as missing beans
spring-petclinic	Boot 4.1, controllers + repositories	0 of 6	0	0 (every edge is a Spring Data repository)
spring-petclinic-rest	Boot 4.1, service layer, profiles, MapStruct	21 of 36	8	1: @Profile alternatives reported as ambiguous
Maven + Lombok	Maven + Lombok	3 of 3	0	1: parameter-name bean selection (-parameters) reported as ambiguous

All three false positives were fixed, and each has a regression test. A fourth class (@Named parameters being ignored) was found and fixed during the qualifier work in Phase 7.

Comparison with real Spring

Each seeded mistake was compiled with SpringGraph and also started in a real Spring context:

Mistake	SpringGraph	Spring
Missing implementation	SG001 at compile time	NoSuchBeanDefinitionException at startup
Second implementation	SG007	NoUniqueBeanDefinitionException
Constructor cycle	SG003, with the cycle path, at compile time	BeanCurrentlyInCreationException (root exception, without the path) at startup
Field @Qualifier not copied by Lombok	SG007	NoUniqueBeanDefinitionException

The Spring column is the root exception from a plain Spring context, not Spring Boot's startup failure analysis. Boot's analysis is more readable and, for cycles, does list the beans involved, but only once the application starts.

These scenarios now run automatically in SpringAgreementTest on every build (21 scenarios).

Feedback time

On the pilot, a seeded mistake was reported in about 3 s (compile) against about 11 s (compile plus a Boot context test). The gap grows with application startup time and shrinks if a fast slice test already exists. See developer experience.

Build overhead

Project	What was measured	Median difference
examples/pilot-boot (7 files)	Clean Gradle compile	about 60 to 180 ms, within run-to-run noise
spring-petclinic	Clean Gradle compile	about 14 ms, within run-to-run noise
spring-petclinic-rest (about 190 files)	Annotation processing only (javac -proc:only), not the project's own build	about 400 ms; run-to-run variation was about 4 s

The spring-petclinic-rest number is processing time, not a full-build measurement. Build overhead was not measured on a large application, and incremental builds were not measured. See performance.

What the pilots showed is still missing

• Low coverage on infrastructure-heavy apps. Most edges are repositories or Boot beans, which are never counted as validated.
• Beans from the app's own @Configuration still need boundary declarations. Limited @Bean analysis (Phase 7) would remove them.
• Profile-switched implementations need one boundary declaration per interface.