Read the reference, not just the citation
An answer can sound precise while leaving an important assumption unstated. Open the referenced passage and compare its actual scope with the claim. Does it describe the same material, conditions, measurement, or population?
Document extraction is imperfect. If a table, equation, or figure matters, inspect the original PDF. A reference makes a claim easier to review; it does not make that review unnecessary.
Follow a result back to its inputs
For a numerical result, inspect the dataset, split, configuration, and saved output. These are part of the explanation. A short answer can be useful at the surface while the supporting record preserves the detail.
Use the project’s provenance to keep derived work connected. When inputs or assumptions change, preserve the distinction between the old experiment and the new one.
Leave a useful record
Unsuccessful attempts are worth keeping when they explain what was tested and why the next step changed. Record the conditions and limits of the lesson so that a later investigation can assess whether it applies.
When you export, capture settled work as a fixed report snapshot. The goal is a record another researcher can inspect, question, and build on.