Evidence for researchers

Your first question: Will an analysis on the synthetic rows give the same answer, with honest uncertainty, as on the real data, and can I reproduce the file?

Short answer: the validity part isn't measured yet. The research-validity report (roadmap #7) is planned, not built, so Ground can't yet tell you whether a regression coefficient or a confidence interval computed on synthetic rows matches the real one. What it can show is how the data is made and measured, where each file came from, and that the same seed gives the same file.

Research validity: pending

What the research-validity report will measure
MeasureResult
Estimates and confidence intervals, real vs synthetic (how often the synthetic interval covers the real estimate)pending first independent run
Effective sample size: how many real observations the synthetic rows are worthpending first independent run
Utility of real plus synthetic rows against real rows alonepending first independent run

Methods

  • Synthesizer. For an upload, several engines (copula and forest-partitioned mixtures, sequential trees, a causal-graph engine, flow matching, and optional GPU engines) are fitted on 90% of the training rows and scored on the rest by the same report users see; the best one wins only if it beats the copula by a point.
  • Report. Column shapes, column pairs and a real-vs-synthetic classifier make the Overall score; utility is train-on-synthetic, test-on-real; privacy is four membership attacks, distance to closest record, attribute inference, singling out and a canary audit. Definitions and limits are in how to evaluate synthetic data quality.
  • Assessments. Each new technique is prototyped behind an option and measured on held-back data before it ships. Two were measured no-go and aren't offered: failure-driven generation and a difficulty rubric.
  • Known limits. Values stay within the range seen in the source; very fine interactions can be smoothed; the data keeps the source's biases unless you rebalance; default-mode privacy is measured, not guaranteed.

The full technical reference is docs/methods.md in Ground's repository, which isn't public yet.

Provenance

Every export gets a signed manifest next to the file: the dataset, model version and engine, settings, privacy guarantee, watermark columns and the file's SHA-256, signed by Ground. The API's provenance check confirms Ground issued a manifest and that it hasn't changed. Extensions of uploaded data also carry a statistical watermark in a few number columns, checkable through the API; it survives reordering and dropping rows, but not rounding the marked columns.

Same seed, same file

For designed datasets without AI-written text, generation is a pure function of the specification, the seed and the engine version: the same three give the same rows, whichever way an export is split into chunks. The worked examples publish the proof: download a file and compare its checksum, or regenerate it. AI-written text comes from a language model and isn't covered by this guarantee, and neither are models fitted to uploads, which this check doesn't test.

Worked examples: seed and SHA-256 of each published file
Example, fileSeedSHA-256
Generate an e-commerce test dataset, data.csv991d08df533c440dc988b5637afeedd70f84f86aa7714c7492fbbfc5679e39b4772
Generate an e-commerce test dataset, data.jsonl9919440833e5924ee5b4cda8ae124f8a10a755965443f8a8b45450dbe5faa2682d9
Generate an e-commerce test dataset, spec.json9912940bc363b451eafe6b9f64417f3bc31d2080bd7a6a3fc86ef64cd173bfe2a6f
Generate transactions with rare fraud labels, data.csv31337ef859b56f8c95f1cc8fe9facbe030209e4266370d82b22f9a7d12f621495dd05
Generate transactions with rare fraud labels, data.jsonl313372a7b292e3f343193f774a94c1a57601b8133fe540c4431bc8fdbf42375459aa2
Generate transactions with rare fraud labels, spec.json313379691703521f68aca8d940db536d1c73fc9b51a59acd4d368fe83b68f139ce3e1
How it is checked
The example files are regenerated and compared byte for byte (npm run examples:generate -- --check); each manifest records the seed and a fingerprint of the engine source. Tests pin output hashes for every template, and check that rows don't change when an export is cut into chunks or when an unrelated column is added.
Limit
A future engine version can change generated values for the same seed; keep the downloaded file when you need exact long-term output.
Benchmarks
The benchmark run records its split and seeds but not yet its commit, machine and package versions, and its harness is in Ground's private repository, so it isn't independently reproducible yet. See the benchmark walkthrough.

Raw outputs