Trust & Responsible AI

Statistical Grounding

All AI personas are grounded in official government statistics from 24 countries (KOSIS, US BLS, Japan e-Stat, and more). This prevents fabricated demographic distributions and ensures personas reflect actual market realities.

3-Layer Voice Sanitizer

Every LLM output passes through three independent quality checks:

• Locale filter — prevents cross-cultural voice leakage (e.g., Japanese persona using Korean expressions)
• Brand-name bias detection — flags spontaneous brand mentions for review against a 28-brand reference dictionary
• Auto-translation pattern detection — identifies machine-translated tone and triggers persona regeneration

Multi-Model Cross-Runs

Every tier we sell, from Hypothesis to Triangulated, runs independent inference across three model families (Anthropic Claude, OpenAI GPT, DeepSeek) in round-robin; Triangulated Pro uses Google Gemini in place of DeepSeek. This reduces single-vendor bias and improves output consistency.

Slot-Based Diversity Enforcement

Occupation and demographic distributions are pre-defined in slot pools across 8 categories. This prevents LLM-induced clustering and ensures statistical diversity in synthetic respondent samples.

Workspace Data Isolation

• Row-Level Security (RLS) on all customer data
• The persona pool is shared across workspaces (personas contain no customer input); customer inputs and results are isolated per workspace
• Read-only sharing links that expire after 30 days by default (up to 1 year)
• Customer data is never used to train models