VECTOR: Achieving Cloud-Scale Repository Modification with Local 7B Models via Deterministic Verification

Description

VECTOR is a research preprint introducing Task-Scoped Deterministic Context (TSDC), a Code Property Graph-based compression system that reduces 50,000+ token repositories to ≤2,500 tokens of structured, call-graph-aware context — enabling a local 7B model (Qwen 2.5 Coder) to perform repository-level code modification with zero cloud dependency and zero fine-tuning. The paper quantifies the 66-point gap between HumanEval performance (84.8%) and real repository modification performance (18.2% pass@5 on Flask), and introduces a 5-layer deterministic verification pipeline that prevents hallucinated code from ever reaching the file system. Full results, benchmark suite (83 tasks across Flask and FastAPI), and VS Code extension (116+ installs) are open-sourced.

Authors

DOI: 10.5281/zenodo.20708135

Publication Date: 2026-06-15

Back to publications list


About