No universal stars
A single number hides task fit, environment, trade-offs, and failure modes.
The test method
A verdict you can inspect.
The method makes disagreement useful. You should be able to rerun the work, find the assumption, and see where your result diverges.
- 01 / Declare
Name the surface.
Record the tool, version, release channel, environment, date, region, plan or price context, task, input, and expected result. Separate vendor claims from the test hypothesis.
- 02 / Run
Keep the receipts.
Capture commands, configuration, prompts where relevant, elapsed time, direct cost, retries, outputs, and observable behavior. Redact secrets and personal data before publication.
- 03 / Break
Leave the happy path.
Test a relevant constraint, malformed input, partial failure, timeout, cancellation, permission boundary, or recovery path. The point is to find the edge of the claim.
- 04 / Review
Separate fact from judgment.
Publish observed facts, supported interpretation, and the recommendation as distinct parts. Name the editor, technical reviewer, conflicts, review copies, affiliate relationships, common ownership, and limitations beside the verdict. When no independent reviewer exists, say so plainly instead of implying one.
- 05 / Reproduce
Make divergence visible.
Provide the smallest safe fixture and instructions another developer can run. If the real system cannot be shared, document the missing layer and reduce the strength of the claim.
- 06 / Maintain
Date everything.
A tool test is a snapshot, not a permanent ranking. Material product changes trigger a retest, a visible stale notice, or retirement. Corrections preserve the original claim and explain what changed.
Measurements stay bounded to the job.
No purchased conclusion
Sponsorship, access, affiliate links, and common ownership are disclosed and never determine the result.
No invisible correction
Material edits receive a date and reason. A challenged result is checked against the saved rig.
No permanent ranking
A versioned product can outgrow a test. Stale evidence is labeled, rerun, or retired.