We need some data to test the actual matching algorithm, i.e. the computation of ExposureInformation and ExposureSummary objects from diagnosis keys and their associated bluetooth payloads / metadata.
More precisely we need:
a set of diagnosis keys, each with a set of captured data (doesn't matter if the asociated metadata is encrypted or decrypted)
correct ExposureInformation data computed from these data sets (probably multiple ExposureInformation items)
data for an ExposureSummary item computed from these data sets
It would be awesome to have ExposureInformation/Summary items computed with the actual EN framework to be sure that everything is correct.
If that is not possible, we could use the microg implementation to generate ExposureInformation/Summary items from a set of manually created diagnosis keys and bluetooth payloads. Thus we could at least verify that the CoraLibre and microg implementations yield the same results.
We need some data to test the actual matching algorithm, i.e. the computation of ExposureInformation and ExposureSummary objects from diagnosis keys and their associated bluetooth payloads / metadata.
More precisely we need:
It would be awesome to have ExposureInformation/Summary items computed with the actual EN framework to be sure that everything is correct.
If that is not possible, we could use the microg implementation to generate ExposureInformation/Summary items from a set of manually created diagnosis keys and bluetooth payloads. Thus we could at least verify that the CoraLibre and microg implementations yield the same results.