Original editorial revision: 9 October 2026. Expanded edition: 14 October 2026.
Research design and arithmetic, not model recommendations. Classification scores are not monetary returns; a registered artifact and a passing software test are not independent strategy validation.
A metric needs a population, label definition, period and accounting convention. Report abstentions, missing labels, exclusions and unresolved outcomes. A convenient win rate without those objects is not an inspectable result.
Repeatability requires source identity, settings, environment and data versions. Repeating a calculation on the same inspected sample checks reproducibility, not independent validation. Preserve the execution contract too.
Write what would reject the candidate before selecting parameters. A flexible story that can explain every result is not a bounded hypothesis. Separate software tests, causal-data checks and independent strategy evaluation.
Criteria as rejectable questions
Measurability asks whether the population, labels and accounting are defined. Repeatability asks whether the exact inputs and calculation can be reconstructed. Testability asks what observation would invalidate the claim. These do not form an editorial star rating. A method can be clearly specified and still fail financial evaluation; a result can be reproducible on one inspected sample and still lack independent support. Use criteria to identify the next artifact, not to conceal evidence absence inside a favorable weighted total.
A three-layer example
The default probability packet is measurable because it declares one hundred binary labels and constant group probabilities. It is repeatable because the compact groups and expanded CSV produce the same Brier loss. It is testable because wrong counts, an empty total or varying row predictions would break its declared calculation contract. These properties establish a software teaching example only. They do not establish a trained forecast, independent sample or profitable account. Applying the same words to real research requires the actual source and reserved evaluation, not merely the ability to fill a form.
The missing-evidence branch is a useful result
If a proprietary model omits its feature availability or exact selection process, report that omission narrowly. Do not infer that its operator used leakage, and do not infer that attractive results imply an honest holdout. The criteria can identify what is unresolved without making either unsupported accusation or endorsement. A claim strengthened later needs the new primary evidence tied to its original scope. This publication cannot create a certification from self-entered evidence states. The release workbench separates documented, unresolved and not applicable rather than auto-admitting candidates.
Preserve economic and operational ownership
A trading study needs native entry and exit rules, fill conventions, risk bounds and complete financial accounting. A deployed product additionally needs the actual artifact and receiving route. Neither layer can be replaced by a generic criterion label. Keep a component fixture in the software-correctness record and a complete independent master in the research record. Missing production promotion evidence remains missing after a model document is uploaded. The cards demonstrate how to write these distinctions plainly, even before any real instrument data is available.