Privacy model
Each workload carries a declared privacy model, its parameters and its limitations.
Technology
Configured Privacy, Measured Utility, Reproducible Releases.
Qualimetrix configures a privacy model per workload, measures utility per task, and records both under a release identifier. It does not offer one default setting that fits every dataset.
How a release is produced
Task-level measurement
Structural fidelity, statistical fidelity and downstream performance are reported separately.
Release record
Methods, parameters, results and limitations are attached to a release identifier.
Hybrid cryptography
Transport between zones negotiates classical and post-quantum key establishment together.
Terms used precisely
The vocabulary a reviewer will ask about.
Privacy unitThe entity whose contribution is bounded: usually a person, sometimes an episode of care.
Adjacency definitionThe pair of datasets a formal guarantee is stated over.
Contribution boundsThe cap on how much a single privacy unit can influence a release.
CompositionHow repeated releases from the same source accumulate disclosure risk.
Held-out real evaluationMeasurement against real records that were never used to build the counterpart.
Cumulative consumptionThe budget already spent across accepted releases from one source.
Formal guarantees
ε is not a complete privacy statement.
A parameter quoted without its privacy unit, adjacency definition and contribution bounds cannot be compared across systems. Qualimetrix states all four, and states which of them a workload does not use.
Transport
Hybrid while the migration runs.
X25519 and ML-KEM-768 are negotiated together, so a failed post-quantum negotiation leaves the session classically protected. SLH-DSA signing of long-lived artefacts is in progress and is not available in a supported release.
Boundary
Formal privacy definitions and supplementary tests may be used. Testing cannot prove the absence of every future attack, and complete behavioural replication of an enterprise system is not a current product claim.
Before acceptance
Five controls before acceptance.
Scope
Workload, data boundary and the decision the evidence must support
Profile
Schema, distributions and constraints read in place
Generate
Counterparts produced against the declared privacy model
Measure
Structural, statistical and downstream results reported separately
Accept
A named owner accepts the record, or the release does not ship
A step is not complete until its artefact exists in writing.
