Research

We publish our methods. We open our architectures.

Our work is built in the open. We believe the fastest way to advance synthetic human intelligence is to share what we learn because breakthroughs compound when they’re shared.

Published Papers

February 2026 · zeldaLabs

Layered Persona Synthesis and Adaptive Social Simulation via Large Language Models

Introducing PersonaGen, a six-layer cognitive architecture for generating high-fidelity synthetic personas that think, remember, and evolve across simulations. The paper details our approach to psychometric grounding, memory systems, and multi-model orchestration.

105+
Psychometric dimensions
96.5%
Coherence pass rate
6
Generation layers
9+
LLM providers

Lab Notes

Short dispatches from the frontier.

NewTownFolk Study|July 2026|7 min

We ran Meta's WhatsApp Plus launch past 1,000 Indian consumers before the market did

Meta priced WhatsApp Plus at ₹79 a month for India. We put the launch to an India-weighted population of 1,000 synthetic consumers before the market answered. Not one said it would definitely subscribe, and the wealthiest refused at almost exactly the rate of the poorest. A rehearsal gives you the trend, not the market share.

Research|May 2026|5 min

A population that has something to lose

Our synthetic personas could reason about almost anything. The one thing they couldn't feel was the weight of a price. So we gave them an economy: zeldaCoins, a working currency that turns stated preference into revealed preference.

Healthcare|April 2026|8 min

Healthcare has always been deeply human. The decisions shaping it deserve to be too.

Focus groups won't show you how a refill reminder lands at 7pm on a Tuesday. We run your messaging against thousands of synthetic patients, HCPs, and payers before a single real person sees it.

April 2026|6 min

Why single-agent AI can't model disagreement

Asking one LLM to play both sides of a debate produces the illusion of disagreement without the substance. Here is what we found when we split the cognitive load across 9+ models.

May 2026|7 min

What happens when you give 1,000 agents moral foundations

We generated a population of 1,000 personas with full Moral Foundations profiles and ran a policy simulation. The emergent patterns surprised us.

March 2026|6 min

The validation problem in generative simulation

How do you know if your synthetic population is realistic? Our three layer approach to evaluating persona quality against published psychometric norms.

February 2026|6 min

Multi-model cognitive diversity: why we use Multi-LLMs

The architectural decision to back different personas with Claude, GPT 4o, Gemini, Grok, and others. What happens when disagreement comes from the models themselves.

February 2026|7 min

Synthetic consensus is not real consensus

The epistemological challenge of drawing conclusions from AI generated debates. What synthetic simulation can and cannot tell you.

Stay in the loop

Sign up for Lab Notes by zeldaLabs

Short dispatches from the frontier of synthetic human intelligence. No spam. Unsubscribe anytime.