Science · Validation
A methodology is only as valuable as its control mechanisms.
A result is not declared reliable on the basis of a single indicator. We distinguish several controls: external reference, deviation measurement, sensitivity to assumptions, reproducibility and auditability.
Reference. Deviation. Variation. Audit.
- 01Reference: typologies are compared with public qualitative studies.
- 02Deviation: the statistical coherence of populations is measured against public benchmarks.
- 03Variation: robustness is tested through parametric variations.
- 04Audit: the methodological file can be examined by an independent third party.
Measure what deviates.
Synthetic populations are evaluated on their deviations from target distributions at three levels and documented in a calibration report appended to the research file. Deviations are read separately: no overall reliability score is derived from them.
- Measure
Univariate marginals
Fit of each attribute considered in isolation: a minimum condition, insufficient on its own.
- Measure
Binary distributions
Fit of two-attribute cross-distributions: the first test of correlations.
- Measure
Ternary distributions
Fit of three-attribute cross-distributions: this is where methods begin to differ.
- Document
Calibration report
Set of documented deviations for a simulation, verifiable by a third party.
SourceStandardised public benchmarks (NPORS); calibration reports appended to the investigation file
Divergence measures are documented by simulation and published on benchmarks reproducible by third parties.
Vary the assumptions.
Structural assumptions — calibration constraints, typologies, interview protocols — are modified within reasonable bounds and then re-simulated to observe whether the conclusion remains stable.
Conceptual illustration: no numerical data are shown. A robust result remains clustered; a fragile result disperses.
A robust result remains probabilistic. It persists under variations without becoming a deterministic prediction or being immune to exogenous events. An insight that does not withstand a reasonable variation is presented as such.
Being able to trace the conditions that produced the result.
- 01Interview protocol and simulation configuration.
- 02Synthetic population and selected typologies.
- 03Reference data and calibration constraints.
- 04Structural assumptions made explicit.
- 05Variations tested and sensitivity measured.
- 06Result delivered and linked to the interviews that produced it.
Auditable ≠ audited. The file can be made available to an audit firm, an independent academic reviewer or an ethics committee; this does not constitute a completed audit or a certification. Published protocols allow a third party to reproduce or challenge the results.
Predictive AI for fraud detection.
What was checked
- Cross-reference against public studies on inequalities in access to healthcare.
- Eleven configurations tested, with deviations documented quantitatively.
- Traceability to the interviews that produced the result, made available to the ethics committee.
What validation does not prove.
Limitation
Robustness is not certainty.
- Projections are probabilistic: they reduce uncertainty before a decision; they do not eliminate it.
Limitation
Statistical coherence is not behavioral validity.
- Measured deviations concern distributions, never the correctness of an individual reaction.
Limitation
Validation is not automated decision-making.
- No overall reliability score, no label: the decision remains human.
Your next decision
Which decision do you want to explore?
Describe your need. We can point you to the right level of support.
