Section 01
The Gap Between Demos and Decisions
Simile raised $200M at a $2B valuation to build simulation infrastructure for enterprises — CVS Health models patient journeys with their platform, and their confidence research sets the technical bar for the field. That's the right scale for Fortune 100 deployments.
But the founder who wants to know whether nurses or administrators buy their scheduling tool doesn't need 8 billion digital twins. They need a fast, honest, affordable simulation of their actual market — with error bars they can see. That's the product we're building, and this post is a transparent tour of what's under the hood in 2026.
Section 02
The Focus-Group Engine: 7 Phases of Moderated Discussion
Our biggest simulation upgrade this year was the group conversation engine. Instead of personas answering in isolation, they now participate in moderated multi-phase discussions — scenario reaction, objection surfacing, social influence, and final reflection.
- Psychological speaker selection. Who talks next is weighted by Big Five-derived talkativeness, dominance, reaction latency, and social-graph reciprocity — not round-robin.
- Realistic pass behavior. Quiet personas skip turns (~10–20% depending on talkativeness), the way real focus-group participants disengage.
- Per-persona temperature calibration. Each persona's LLM temperature derives from their activity profile — consistent characters, not uniform randomness.
- Parallel execution. Group simulations now run concurrently across clusters, cutting wall-clock time dramatically on large panels.
The output isn't just a transcript. Every conversation produces per-persona summaries, an influence graph with Cialdini-style persuasion mechanisms, and herd-cascade detection — the moments where one persona's opinion visibly shifted others.
Input
- Setup
- 8 personas, post-survey focus group
- Phases
- 7 (scenario reaction → final reflection)
- Turn logic
- Psych-weighted selection + pass/skip behavior
- Runtime
- Parallel across clusters
Output
Illustrative structure of a focus-group run. Every turn names its target (@FirstName), reactions are activity-weighted, and each persona's summary is stored for the network visualizer.
Section 03
Multi-Model Inference: The Right LLM for Each Job
One model doesn't fit every job in a simulation pipeline. We route each step to the model that does it best — and this year we rebuilt that routing layer:
| Pipeline step | Model | Why |
|---|---|---|
| Persona responses | Kimi K2.5 (BaseTen) | Best persona-consistency in our tests |
| Simulation setup & situations | GLM-4.7 | Strong structured generation |
| Fast trigger turns | GPT-OSS-120B (Cerebras) | 3,000+ tokens/sec for real-time feel |
| Post-survey conversations | DeepSeek-V4-Flash (BaseTen) | Cost-optimized at conversation volume |
| JSON repair | GPT-OSS-120B | Reliable structured output fixing |
The routing matters more than any single model choice. A 200-persona survey makes thousands of LLM calls; mixing fast models for interactive steps and cheap models for bulk conversation generation is how real-time simulation stays affordable.
Section 04
Seeing the Market: The Network Visualizer
This year we rebuilt the response network visualizer around a simple thesis: spatial position carries information that averages destroy. A 3.2 average rating can hide a segment in revolt.
- Force-directed clustering. Personas self-organize by sentiment and similarity — tight red clusters reveal segments in revolt that bar charts average away.
- Live streaming. Responses appear as nodes in real time via Convex subscriptions as the LLM pipeline generates them.
- Persona detail cards. Click any node for demographics, reasoning, and their conversation summaries from the focus-group engine.
- Dark-theme graph layouts across overview and report tabs, with avatar-backed nodes so the market looks like people, not dots.
The clustering layer got particular attention: sparse clustering improvements mean segments separate cleanly instead of forming one gray blob — the difference between a graph you can read and a graph that's decoration.
Section 05
The Confidence Layer: Stability Verification
The most important ship of the year isn't a feature — it's epistemic honesty at scale. Every survey now computes a stability score using split-half TVD: we randomly split the panel in two, 20 times, and measure how much the halves agree.
Stable panels this month achieved split-half TVD as low as 0.05 — random halves of the panel agree within 5 points. Full methodology in our confidence post.
The badge on every survey detail page shows the bucket, the TVD value, and flags individual questions with missing ratings, degenerate distributions, or low-quality reasoning. It's the same Total Variation Distance math from Simile's confidence research, applied to internal consistency — an honest lower bound while our benchmark program builds toward calibrated accuracy.
Section 06
Where We're Headed
- Benchmark calibration at scale. Growing from 35 Pew-anchored questions to 200+, converting stability scores into calibrated accuracy predictions.
- Cross-model TVD. Re-running a sample through a second LLM to detect model-specific bias — the slot is reserved in the confidence schema.
- Deeper world-state simulation. Personas that react to breaking news and market events, not just static surveys.
- Longitudinal persona memory. Panels that remember prior studies and develop genuine research-informed opinions over time.
“Simile is proving simulation matters at enterprise scale. Our job is proving it works at founder scale — with published error rates, honest limitations, and a price that lets you test ten ideas instead of one.”
FAQ