Agentic AI, Software Engineering
Sphragis
A pre-registered study of what coding agents learn from code review
A coding agent becomes useful once it knows how an organization writes software: its internal APIs, error-handling patterns, naming rules, and what its reviewers reject. Sphragis tests whether that house style can be learned at all from an organization's public code-review history, and what a declared boundary protects once it can. The study is pre-registered: its pass rule is fixed in code and its test window is sealed before collection. It is written as a Registered Report for MSR 2027.
Built with Python, PyTorch, Transformers, PEFT
Contributors
AJ BareaRIT PhD Researcher · Sphragis · Pharos · Phalanx-FL