EVIDENCE-GUIDED LOCAL CODE GENERATION AND REFINEMENT: INTEGRATING DETERMINISTIC GUARDRAILS, BOUNDED GRAPH RETRIEVAL, AND DIAGNOSTIC-AWARE EVALUATION
This study investigates a resource-conscious approach to local code generation and refinementthat integrates three complementary mechanisms: Bounded, task-directed graph retrieval forselecting repository context; deterministic guardrails and execution-based validation for checkingproposed changes; and diagnostic-aware evaluation to distinguish final task outcomes fromobservable properties of the development process.
Authors
- Sai Ishaan Balaji (ORCID: https://orcid.org/0009-0000-5717-4050)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-10-31
- DOI
- https://doi.org/10.5281/zenodo.23833482
- Primary Topic
- Software Engineering Research
- Type
- preprint