Behavior tested
The workflow or asset was executed against a documented input, with the relevant result and checks retained or linked.
Does not mean universal accuracy, repeatability, or cross-model performance.Testing methodology
How Grow AI Skills tests what it publishes.
We label the evidence we actually have, preserve the boundary of each test, and show readers what a result does—and does not—support.
Evidence levels
Three labels, without a vague score.
A higher-sounding badge is not useful unless its meaning is inspectable. These labels describe the work completed for a specific asset or workflow.
Structure checked
The package, required fields, file structure, or reusable contract was inspected against defined requirements.
Does not claim that the exact reusable asset was executed end to end.Source reviewed
The instructions, origin, and supporting sources were reviewed, but no first-hand behavior test is claimed.
Readers should treat tool behavior as unconfirmed until it is tested.The workflow
From a bounded claim to a reviewable result.
Every test starts by narrowing the question. A passing result applies only to the named inputs, environment, date, and checks.
- Define the boundaryName the task, tool, version or interface, expected result, and claims that remain out of scope.
- Prepare safe inputsUse authorized, representative, or synthetic material and record any protected facts that must not change.
- Run the workflowFollow the published instructions and retain the output, run record, or other inspectable evidence when practical.
- Check the resultCompare the output with explicit acceptance criteria; use negative or mutation tests when false passes matter.
- Publish the boundaryShow the status, date, sources, reusable materials, known limits, and what the evidence does not prove.
- Retest material changesUpdate the evidence when a tool, interface, source, or workflow change could alter the conclusion.
What we record
Evidence is more than a pass label.
- Source basis and first-hand versus vendor-supplied claims
- Execution environment, named tools, and verification date
- Inputs, outputs, checks, and downloadable review materials
- Reproducibility limits, blocked steps, and untested branches
- Human decisions required before publication or real-world action
Inspect the method
Use the same gates in your own workflow.
Review the editorial rules, download the evidence-gates starter kit, or report a result that should be retested.