Does the data actually hit the target?
The point of declaring an outcome is that the generated rows satisfy it. We measured how well they do. When you ask for revenue that sums to an exact monthly curve, imitation synthesizers trained on the same data miss those targets by 74 to 86 percent. Misata hits them to zero error, because it solves for the outcome in closed form rather than learning an approximation.
| Measure | Imitation synthesizers | Misata |
|---|---|---|
| Declared monthly aggregate error | 74 to 86 percent off target | 0.00, exact by construction |
| Foreign-key integrity across tables | Varies, not guaranteed | 100 percent, orphan count reported |
| Needs real data to start | Yes, trains on a sample | No, generates from specification |
| Deterministic and reproducible | Stochastic training | Same seed, identical bytes |
| Fidelity to a real dataset's joint distribution | Strong, its core strength | Approximate via mimic mode |
How the test works
We declare an aggregate target, for example monthly revenue rising from 50,000 to 200,000 across a year, and ask each method to produce a dataset that meets it. Then we bucket the generated rows by month, sum the measure, and compare against the declared target. Misata solves the per-period allocation in closed form, so each period sums to its target to the cent. Imitation methods can only reproduce patterns present in their training data, so a target the training data did not already reflect is missed by a wide margin.
What this does and does not claim
This is a controllability and integrity benchmark, not a fidelity benchmark. For learning the full joint distribution of an existing real dataset, a trained generative model can go further, and we concede that openly. Misata is built for the other regime: cold-start generation with exact aggregate control and integrity by construction, where imitation methods cannot serve because there is no data to imitate or the totals must hit an exact number. The mechanism and the formal error bound are described in the paper.