Match the question
Define the claim and the population it concerns. A convenient benchmark may not reflect the intended setting.
Separate development and evaluation
Plan splits before repeated experimentation. Consider leakage through related records, time, subjects or preprocessing.
Report uncertainty
Explain variation across runs and meaningful limitations. A single score rarely tells the whole story. Include appropriate baselines and error analysis.
Keep the record
Document decisions and exclusions. Clear reporting helps others understand what your evidence supports and what remains open.