---
title: "Synthetic Research Evidence Center | Minds"
canonical_url: "https://getminds.ai/research/synthetic-research-evidence-center"
last_updated: "2026-08-26T19:33:02.850Z"
meta:
  description: "Inspect Minds validation, method pipelines, audience grounding, legal and procurement documentation, model-change provenance, and the limits of synthetic..."
  "og:description": "Inspect Minds validation, method pipelines, audience grounding, legal and procurement documentation, model-change provenance, and the limits of synthetic..."
  "og:title": "Synthetic Research Evidence Center | Minds"
  "twitter:description": "Inspect Minds validation, method pipelines, audience grounding, legal and procurement documentation, model-change provenance, and the limits of synthetic..."
  "twitter:title": "Synthetic Research Evidence Center | Minds"
---

Minds

August 21, 2026·Methodology·Minds Team

# **Synthetic Research Evidence Center**

Minds treats trust as a chain of inspectable evidence: audience sources, answer-level provenance, versioned research methods, task-specific validation, and public procurement documents.

[Run inspectable research](https://getminds.ai/?register=true)

Synthetic research should not be trusted because an answer sounds human or a vendor publishes one large percentage. Trust requires a chain: the audience definition, the data and assumptions beneath it, the questions and stimuli, the answer-level evidence, the calculation method, the validation reference, and the system version.

Minds calls this approach inspectable synthetic research. It does not eliminate uncertainty or turn synthetic respondents into a probability sample. It makes the evidence and boundaries easier to review before a team acts.

## Evidence map

| Evidence layer | Public artifact | What it supports | What it does not prove |
| --- | --- | --- | --- |
| Audience grounding | [Synthetic Audiences methodology](https://getminds.ai/research/synthetic-audiences-methodology) | How briefs, sources, distributions, and assumptions should be handled | That every audience is representative |
| Applied validation | [Gen Z food survey replication](https://getminds.ai/research/synthetic-gen-z-food-survey-validation-2026) | Aggregate agreement for one outcome-blind instrument | Universal or individual accuracy |
| Method implementation | [Research method pipeline catalog](https://getminds.ai/research/research-method-pipeline-catalog) | Which methods are registered and how stages produce artifacts | That synthetic inputs equal live-human evidence |
| Quality control | [Validation checklist](https://getminds.ai/research/synthetic-audiences-validation-checklist) | A pass/fail review before using a result | Formal validation for every decision |
| Procurement | [DPA](https://getminds.ai/legal/dataprivacy), [subprocessors](https://getminds.ai/legal/subprocessors), [TOM](https://getminds.ai/legal/tom), [SLA](https://getminds.ai/legal/sla), [DPIA](https://getminds.ai/legal/dsfa) | Public processing, security, service, and risk documentation | Customer-specific commercial terms |
| Change provenance | [Model-change policy](https://getminds.ai/research/model-change-validation-provenance) | How material research-system changes affect claims | That every model update is behaviorally neutral |

## Audience construction evidence

A reusable audience should preserve why each segment exists. Minds supports audience creation from existing Minds, explicit descriptions, attached files or links, approved research, partner material, and suitable public sources. Completed reports or respondent datasets can provide observed distributions. Screeners and unfielded questionnaires can define options, exclusions, and target quotas.

Observed distributions, targets, and assumptions are not interchangeable. If a source specifies a variable without a share, a proposed distribution can be labeled as assumed. If only marginal evidence exists, a joint distribution may use an explicitly independence-based reconstruction instead of pretending a correlation was observed.

This source discipline is explained further in [synthetic audience data sources compared](https://getminds.ai/comparison/synthetic-audience-data-sources-compared).

## Answer-level evidence

Aggregate charts are easier to consume, but they can conceal generic, unsupported, or contradictory answers. Supported Minds workflows keep the individual response or transcript available so a researcher can examine what produced the summary. Audience definitions and source context remain part of the study record, and structured exports are available for eligible workflows and plans.

Inspectable does not mean infallible. A source citation shows where context came from; it does not prove that the context caused the answer. A fluent quote remains synthetic and should be labeled as such.

## Versioned method evidence

Minds separates qualitative generation from deterministic analysis. The registered catalog defines method inputs, configuration, pipeline stages, calculator versions, and artifact contracts. Available methods include questionnaires and qualitative exploration plus ranked preferences, segment comparison, MaxDiff, conjoint, NPS, top/bottom-box scoring, key-driver analysis, TURF, Gabor-Granger, Van Westendorp, and Kano.

Conjoint, for example, produces a design, responses, an estimate, diagnostics, simulated shares, and evidence. MaxDiff produces responses, estimates, diagnostics, and ranked evidence. This structure makes a calculation reproducible from its inputs. It does not make the synthetic responses equivalent to observed purchase behavior.

## Applied validation evidence

The current public Minds validation used the original questions and answer options from a UK Food Standards Agency study while keeping the human percentages outside the runtime. A locked audience of 301 persistent Gen Z Minds completed 903 answers. Across 21 answer-option cells, aggregate approximation was 93.99%, defined as 100% minus mean absolute percentage-point error. Mean distribution overlap was 78.98%.

The report includes the design, metrics, source dataset, provenance counts, item-level results, and limitations. It states that the result covers three related questions in one survey and does not establish universal performance. This is the standard Minds intends to apply: a scoped result with a reproducible definition, not an unqualified accuracy slogan.

## Institutional grounding

Minds has a public research and commercial partnership with [SINUS-Institut](https://getminds.ai/newsroom/minds-partners-with-sinus), providing an institutionally grounded audience framework where enabled. Research groups such as [INTEGRAL](https://www.integral.co.at/en/expertise) demonstrate the broader value of combining established segmentation, product-testing, pricing, and online-research practice with modern synthetic workflows; this page does not represent INTEGRAL as a Minds partner.

Partnership grounding is not a substitute for study validation. It strengthens the provenance of the population definition, while the research question, method, and outcome still require their own evidence.

## Procurement evidence

Public legal pages reduce the amount of diligence hidden behind a sales process. Minds publishes its DPA, subprocessors, TOM, SLA, and DPIA. Public [pricing](https://getminds.ai/pricing) gives self-serve buyers a baseline, while the applicable order form governs seats, allowances, enterprise controls, customer-specific onboarding, validation, calibration, integrations, support, and service terms.

Use the [synthetic research procurement checklist](https://getminds.ai/guide/synthetic-research-procurement-checklist) to compare Minds with Electric Twin, Simile, Aaru, Artificial Societies, or another vendor on the same evidence request.

## Evidence boundaries

Minds synthetic outputs should not be presented as real respondent testimony. They do not automatically establish representativeness, causality, precise demand, exact willingness to pay, safety, clinical validity, or legal compliance. The right next evidence step depends on the decision: a live survey, interview, focus group, experiment, behavioral analysis, sales or support review, or expert assessment.

The [validation checklist](https://getminds.ai/research/synthetic-audiences-validation-checklist) should fail a run when the audience is vague, sources are undocumented, prompts are leading, uncertainty is removed, or a synthetic result is presented as final proof.

## Compare the evidence, not the adjectives

Before selecting a platform, compare the same artifacts:

1. audience and source manifest;
2. observed, target, and assumed distributions;
3. raw answers and transcripts;
4. method, calculator, and artifact versions;
5. task-specific benchmark with failures and subgroup error;
6. model-change record;
7. DPA, subprocessors, security measures, SLA, and deletion terms;
8. the required real-world validation step.

For the market view, use the [synthetic respondent comparison hub](https://getminds.ai/blog/synthetic-respondents-comparison-hub), [validation comparison](https://getminds.ai/comparison/synthetic-audience-validation-and-accuracy), and [security and procurement comparison](https://getminds.ai/comparison/synthetic-research-security-and-procurement).

## **Frequently asked questions**

### **What evidence does Minds publish?**

Minds publishes an applied outcome-blind validation, synthetic-audience methodology and checklist, a versioned research-method catalog, public legal and procurement documents, and a model-change provenance commitment.

### **Does Minds claim one universal accuracy rate?**

No. The current 93.99% approximation result applies to one Food Standards Agency survey replication and is reported with its metric, population, instrument, and limitations. It is not individual prediction accuracy or a universal product rate.

### **Can buyers inspect underlying research evidence?**

Supported Minds workflows preserve audience definitions, source context, individual answers or transcripts, method outputs, and exportable artifacts where available. The exact artifacts depend on the workflow and plan.