Why per task
One number hides the failure.
Material fit for load testing is not automatically fit for model training or for prevalence estimates. Each task is measured on its own terms, and a task that was not measured is recorded as not measured.
Evidence
Every Release Ships Its Own Record.
The evidence pack is a first-party technical record: methods, parameters, results, limitations and a release identifier. It is not an independent certification and does not claim to be one.
What is attached
What an accepted release carries.
- Methods and parametersThe privacy model, its parameters and the generation configuration used for this release.
- Utility resultsStructural, statistical and downstream measurements, reported per task rather than as one score.
- LimitationsWhere the counterpart is not fit for a purpose, written into the record rather than left to inference.
- Release identifierA reference that stays inspectable after the release has been used downstream.
- Residual riskWhat remains after the controls described, stated in terms a reviewer can act on.
Specimen
What the record looks like.
Specimen recordRelational and tabularIllustrative, not a customer release
Privacy modelDeclared per workload
UtilityTask-level, held-out
LimitationsStated in the record
StatusSubject to assessment
Boundary
The specimen above is illustrative. Values in a real record are produced by measurement against the workload they describe, and are not transferable to another workload.
Review model
Task-level, not a single score.
Held-out evaluation
Measured against records it never saw.
Downstream results compare the counterpart against real records held out of generation. Degraded groups are named in the record rather than averaged away.
Boundary
A first-party record supports a bounded conclusion. It is not an independent audit, not a certification, and not a warranty of fitness for a purpose Qualimetrix has not measured.
