My tests were green. My code was wrong. The spec was 3,000 years old.

Short summary

Testing in underspecified domains hits the self-validation paradox: your tests encode your code's assumptions, so both pass together even if both are wrong. The author solves this using differential testing—running a large sample of inputs against multiple independent implementations and comparing field-by-field until all agree. With fixed random seeds for reproducibility and separate tests per concern, this pattern generalizes beyond I Ching to any domain where specs are ambiguous.

• •Self-validation paradox: tests written from same misconceptions as code don't catch errors
• •Solution: differential testing against independent oracles (other implementations)
• •Use fixed random seeds and per-concern separation for reproducible, comprehensive validation

Generated with AI, which can make mistakes.