External data: Not allowed unless it is explicitly provided or whitelisted in the “Data” tab.

Pretrained models: Allowed only if trained exclusively on provided/whitelisted data.

Single-account per participant; team mergers per Kaggle policy.

No leakage: Do not engineer features that peek into future timestamps.

Reproducibility: Finalists must provide an inference notebook or script (CPU-only) and the trained model artifact so organizers can run on the validation set.

Determinism: Set seeds; avoid non-deterministic ops that change outputs between runs.