How do you compare age assurance systems?
Compare like with like: fix the object and its version, name the decision the comparison feeds, choose one indicator of confidence and its threshold, and test balanced accuracy, binding, equality and attack resistance with a sample that can support the claim. Publish the method with the result.
Start with the object, not the vendor
A component test and a system test answer different questions. Write down whether you are comparing a face estimation component on a bench, the same component inside a live workflow, or the whole service. Give each object a version and keep it fixed for the run.
Choose the indicator before the test
The framework sets five indicators of confidence. The public working draft of the benchmarking part gives external accuracy benchmarks for four of them, from ninety per cent to ninety-nine point nine nine per cent. Pick the one the decision needs, and write the pass mark down before you see a result.
Measure the things a percentage hides
Report the true positive and true negative rates, the class balance, the error distribution by age band, the near-threshold cases, the equality analysis and the presentation attack result. A single average can be accurate and still fail the children it is meant to protect.
Report the claim with its method
Publish the summary beside the practice statement and keep the detailed report with the limitations. Set the triggers that force a re-run: a new model, a changed threshold, a new attack pattern or publication of the final text.
Not legal advice; does not certify. The draft is a guidance document in ballot. Check the published text before a decision depends on it.