Developer Builds Code-Repair Benchmark That Tests Beyond Prompt Examples
A software developer created 'invariant-repair-suite', a Python benchmark with eight code-repair cases, while participating in the Kaggle Benchmarking Challenge. The benchmark is designed to test whether AI models can preserve code behavior not explicitly demonstrated in the prompt, targeting quiet bugs that pass visible examples but break production invariants. Each case provides a contract, a buggy implementation, and two visible examples, while scoring weights hidden edge cases at 75% and visible regression checks at 25%. Reference fixes scored a perfect 1.0 overall, while the buggy starter implementations scored 0.57, confirming the harness can distinguish correct from plausible-but-incomplete code. The author outlines three core design rules: keep hidden test conditions out of the prompt, penalize regressions, and prefer executable assertions over model-based judgment.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in