01Profile
Map the data-generating process before reproducing it.
Profile entities, sequences, constraints, missingness, rare categories, outliers, and the source fields that require stricter handling.
SchemaDependenciesSensitive fields
Platform
One evidence chain from source data to model decision.
MirrorFoundry connects synthesis, privacy evaluation, task-based utility, and model simulation so teams can release synthetic data with explicit evidence—not intuition.
01
Relational dataTime seriesEvent streamsFree textEdge-case simulationPrivacy evidenceRelational dataTime seriesEvent streamsFree textEdge-case simulationPrivacy evidence
The foundry loopFour stages.
Four stages.
One release record.
Each stage produces evidence that can be reviewed, repeated, and attached to an accepted dataset or model decision.
02Forge
Generate new records from learned patterns and explicit rules.
Combine distribution-aware generation with business constraints so the result preserves useful behaviour without copying source rows.
TabularTextEvents
03Prove
Measure privacy and utility for the intended use.
Run fidelity, task performance, nearest-neighbour, inference, memorisation, and rare-record checks against acceptance thresholds.
FidelityPrivacyCoverage
04Simulate
Stress models across controlled populations and shifts.
Create boundary cases, rare cohorts, and distribution changes, then retain the scenario, model output, evaluation, and reviewer decision.
RobustnessFairnessRelease
Platform layers
Generation is only useful when the evidence travels with it.
A MirrorFoundry implementation includes only the sources, generators, tests, scenarios, and environments agreed in the SOW.
Data profile
Map types, entities, relationships, sequences, missingness, constraints, and sensitivity.
Generation
Forge structured, semi-structured, and text data from learned patterns and explicit rules.
Privacy evaluation
Test disclosure risk, overfitting, memorisation, and similarity to protected source records.
Utility evaluation
Compare distributions, correlations, constraints, and performance on intended downstream tasks.
Model simulation
Create rare conditions, controlled cohorts, and distribution shifts for robustness testing.
Release governance
Retain configurations, evidence, limitations, approvals, and permitted-use conditions.
Start with one release
Define the first accepted synthetic release.
Start with one restricted dataset, one model decision, and explicit acceptance criteria.
Request an assessment