Each of these exists because a study broke without it. They are principles, not a recipe: the specific machinery stays in-house, but you should know exactly what standards your data is being held to.
Structure first, narrative never. Generate and validate structured survey responses against real distributions before writing any persona backstory. Otherwise the model's cultural assumptions quietly drive the data.
Anchor distributions empirically. Key distributions get pre-assigned from real benchmark data or population priors. The model is never allowed to invent the shape of an answer.
Validate at the point of creation. Every persona passes through automated logic checks that route it to pass, retry, or manual review. Bad personas get caught when they're made, not after they've contaminated the results.
Calibrate to the community's real voice. Language models default to agreeable and enthusiastic. Real communities are ironic, tired, or blunt. Every community study ships with an explicit cultural-tone instruction.
Triangulate sources. A single grounding source shapes every persona the same way, and you can't see it from inside. Multiple independent anchors, always.
Disclose everything. Treat divergence as a finding. Every limitation gets published. When synthetic output disagrees with a benchmark, that's investigated, not buried: sometimes it's a failure, sometimes it's a signal the benchmark hasn't caught up to.
The fundamental rule. Trust synthetic data for the direction of a relationship. Never trust it for exact numbers. Use it to pre-test, de-risk, and find patterns. Then validate with real fieldwork.
"These same seven standards are the checklist I run your data against. If your vendor can't tell you how their panel would score on them, that's worth knowing."
One dataset, one week, one plain-language memo on where it holds and where it doesn't. No retainer.