Confidence Techniques for Guiding AI Migration

Migrating a legacy Java codebase (~10,000 lines) to Rust using AI tools surfaces key challenges around code quality confidence. Three practical techniques are presented: (1) adding documentation comments to generated tests that link back to their Java equivalents for traceability, (2) being explicit about architectural expectations upfront to prevent LLMs from inventing inefficient solutions (e.g., using stdio instead of JNI), and (3) using mutation testing tools like cargo-mutants to validate test quality beyond raw coverage percentages. The core insight is that AI makes code generation cheap but shifts the burden to verification — better specifications and robust testing practices are essential to achieve real productivity gains.