Origin Part 19: The Number Was Wrong

Short summary

An ML engineer discovered their model's 88% eval score was inflated by train/test data leakage — 23 of 26 held-out probes were also in training data. After fixing the data generators to guarantee zero overlap and adding verification tools, the honest score dropped to 58%. The author also found that a file labeled 'sandbox' was actually the production checkpoint, revealing how naming discipline gaps caused systemic sloppiness in the training pipeline.

• •88% eval score was illusory — 23/26 held-out probes overlapped with training data due to a race condition in data generators
• •After fixing overlap, honest generalization score dropped to 58% (still 16x above baseline)
• •'Sandbox' file label on production checkpoint caused systemic discipline failures; renamed and added overwrite guards and preflight checks

Generated with AI, which can make mistakes.