Verification-Gated Local Model Cascading

This research release evaluates verification-gated local model cascading for coding tasks through the Adaptive Workflow benchmark. It compares Small-only, Strong-only, and Adaptive conditions across 542 EvalPlus tasks: 164 HumanEval+ and 378 MBPP+. Each completion is evaluated with deterministic base and Plus tests. The Adaptive condition uses a fixed local FAST-to-BALANCED-to-BEST escalation policy with Qwen2.5-Coder tiers, resource monitoring, checkpointed execution, and sanitized reproducibility artifacts. Research status: Corrected preprint and reproducibility release; not peer reviewed. The reported results are bounded to this benchmark, model configuration, and operating point. They do not establish an independently deployable verifier, predictive-routing superiority, equal-budget superiority, or universal efficiency advantage. Plus-test results are Adaptive 331/542, Strong-only 297/542, and Small-only 220/542. Generative AI assistance disclosure: OpenAI ChatGPT/Codex assisted with software implementation, organization, editing, and formatting. The author directed the research question and experimental design, reviewed the final release, and accepts responsibility for the claims and conclusions. Reproducibility code and sanitized results:https://github.com/Mystrowin/adaptive-workflow-research