Verification checklist

Use this for results you will rely on. For the first exercise, start with the expected answer and a rerun. Add checks as the analysis grows.

Copy this into your project as VERIFICATION.md. Link to evidence beside each relevant item: test output, a diagnostic figure or a run record. Mark an item "not applicable" with a reason when needed. See page 7.

Scientific specification

• The question, unit of analysis, units and missing-data policy are documented.
• Model equations and conventions are specified where applicable.
• The implementation matches the intended method.
• Exploratory changes are distinguished from preregistered analyses.

Calculation checks

• A small example agrees with a calculation made independently.
• A deliberately introduced plausible error is detected.
• Tests cover relevant invalid inputs, duplicates and missing values.
• Expected test values were reviewed and not changed merely to pass.
• Diagnostic plots have been inspected and unexpected patterns investigated.

Model fitting, if applicable

• Recovery is assessed over relevant parameters, sample sizes and several seeds.
• Tolerances are justified and weakly identifiable parameters are discussed.
• Convergence and boundary estimates are checked.
• Reported uncertainty is evaluated, including interval coverage where appropriate.

Review and history

• I can explain the key calculation and the limits of its checks.
• Code and new files have been reviewed.
• Any fresh-session review findings have been checked and addressed.
• Small, descriptive commits record the changes and relevant check outcomes.

Reproduction

• A documented command reproduces the result from saved inputs and settings.
• The environment is recorded and the lockfile committed where used.
• Random seeds are recorded where applicable.
• The run record identifies code version, uncommitted changes, inputs and settings.
• A clean rerun produces the expected result.
• Fictional observations are labelled and cannot be mistaken for collected data.
• The AI-assistance disclosure accurately describes what was done.

Cluster runs, if applicable

• Expected outputs are present and contain the required results.
• Missing, failed and timed-out jobs are accounted for.
• Computation ran on allocated compute nodes.

Outstanding questions

Record remaining uncertainties and who will help resolve them. Completion is a judgement about relevant evidence, not a score based on the number of ticked boxes.