Transcript Testing: Writing AI-Generated Code Tests for Human Review

Andrew proposes Transcript Testing, a testing style that records a human-readable transcript of system behavior and asserts it against an expected transcript, to help reviewers verify AI-written code. He illustrates it with an Apache DataFusion PR on Parquet Bloom filter pruning, where a 400-line AI-generated test was replaced by transcript output showing which filters and pages were read. He notes the approach is not new, citing Oracle's ORATEST and DataFusion's sqllogictests, but argues AI now makes transcript generators cheap enough to use far more often.