METHOD · VERSION 1
How Evidence works, and where it stops.
Evidence helps make hypotheses, controls and gaps explicit before field research, professional diagnosis or operational expansion. It does not measure the market, does not represent respondents and does not prove causality.
1. Decision before collection
Every piece of work starts with the decision: the choice, what it can change and who is accountable for it. Only then come the question, the proposed population, the method and the sources. The plan appears before any answer.
2. Evidence states
Every data point carries a declared state, and no state is promoted to another.
3. Synthetic data and reproducibility
When Evidence generates data, it uses a stable seed and deterministic pseudo-randomness, with a declared engine version. The same seed, parameters and version produce the same output. Changes to the hypothesis, parameters, seed or version create a new identifier.
Synthetic numbers have construct validity internal to the model, useful to discuss choices, but they do not estimate a real population.
4. Sources and transparency
External sources give context to the hypothesis. They do not enter the calculation, do not calibrate the model and do not turn synthetic output into a benchmark. The transparency standard follows the need to make sponsorship, population, method, size, processing and limitations inspectable, as in the AAPOR Disclosure Standards. Risk dimensions relate to the NIST AI RMF, without claiming compliance, and data quality practices to the OECD Quality Framework.
5. The decision envelope
Every plan ends in an envelope with what the decision holds, what it does not hold, what is unknown, the current limit, what would change the decision and the next evidence. The envelope is not personalized professional advice, a forecast or a research result.
6. How the Studio generates and measures
Generation. Every column follows a declared distribution (normal, lognormal, uniform, probability, categories, sequence or email on example.com). Correlations between numeric columns use a Gaussian copula: normals correlated by a Cholesky decomposition become each marginal through its inverse. Bounds clamp the value at the minimum or maximum. The seed makes the file reproducible, and the SHA-256 hash identifies each export.
Validity (should be 100%). Types, bounds, integers, allowed categories, unique identifiers, reserved domain in emails and null rate within statistical tolerance.
Fit to the specification. Numeric: 1 minus the Kolmogorov-Smirnov statistic against the declared distribution, accounting for bounds. Categories and booleans: 1 minus the total variation between declared and generated shares. Correlations: the copula target becomes an expected Spearman (6/π · arcsin(r/2)), compared to the obtained Spearman. The overall score is the average. It is fidelity to the specification, not to real data.
Basic privacy. Direct identifiers are generated by the Studio itself. For the marked quasi-identifiers, rows are grouped (numeric ones in quintiles) and we count how many fall in groups with fewer than 5 records. Categories under 1% show up as rare. These checks support design review; they are not an anonymization opinion, and distance to the closest record, on its own, does not prove privacy.
Pasted schema. A CREATE TABLE or a JSON Table Schema becomes editor columns: type, NOT NULL, CHECK (ranges and lists), enum, key and uniqueness. Where the schema does not state a distribution, the Studio uses a default and shows a warning; free text, UUIDs and anything that looks personal become fictitious identifiers. Dates are uniform between start and end, measured by KS, and enter k-anonymity by year.
Formats and seeds. CSV, Parquet (uncompressed, with the label and seed in the metadata) and JSON come from the same dataset; the manifest carries the SHA-256 of each. The seed comparison generates the same design with 5 neighbouring seeds and shows the range: a difference smaller than that is sampling luck.
7. Limits
- There are no respondents, personal data, customer data, probability samples or weighting on the public surfaces.
- There is no competitive comparison, market ranking, causal estimation or forecasting.
- To move to real research: state the hypothesis, approve instrument and population, obtain consent, define collection, preserve provenance, review biases and publish limitations before any external claim.