Test a bounded task
For a calculator, compare outputs with independent hand calculations. For generated strategy code, inspect data access and test the rule on a small constructed sequence. For a forecast, define the target and hold-out procedure before evaluating predictions.
Respect information timing
Document when every feature or document became available. A historical test using a current language model may be affected by knowledge acquired during pretraining. A date in a prompt does not itself demonstrate that later information was excluded.
Preserve an audit trail
Save prompts, model identifiers when available, generated artifacts, data versions and evaluation rules. Separate deterministic calculations from model commentary. Re-running an experiment may still produce different text; record this limitation rather than presenting it as identical reproduction.
Sources & references
- scikit-learn — Data leakage and common pitfalls ↗
- scikit-learn — Cross-validation and time series evaluation ↗
References support specific definitions or methodological context. Examples and research checklists are explanatory; no performance result is claimed.