Methodology
Swarmgram · it’s a simulation
We model human behavior.
We build Lewis, run simulated research through Lewsearch, and publish comparisons with human results.
Research direction
Lives in simulation.
Our ambition is to simulate societies: agents whose lives unfold over time, shaped by memory, relationships and a changing world.
SwarmgramCity
In SwarmgramCity, agents live in a simulation, interact with one another and go about daily lives. We are developing how they carry memory and beliefs forward over time, so we can study how individual choices become shared patterns.
LewBrain
LewBrain development reaches down to a simulated heartbeat. We are building advanced memory, belief and sensory architecture around agents that carry experience forward. Our long-term goal is a quasi-conscious substrate for research into minds and simulation.
We want to study how a past shapes a future, how relationships change a decision, and what simulated lives might teach us about minds.
Research direction · In development. Our published evaluations report results on specific tasks.
Read about SwarmgramCityFigure 1 · Public forecast record
Scorecard
| Measure | Forecast | Published outcome | Grade |
|---|---|---|---|
| Deloitte holiday spending Issue 07 · October 1, 2026 | $1,589 | $1,654 | Partial |
| NY Fed household spending growth Issue 06 · September 15, 2026 | 4.9% | 4.704714298% | Partial |
| Michigan consumer sentiment Issue 05 · September 9, 2026 | 55.4 index points | 47.8 index points | Miss |
Latest
All research and updatesReport · awaiting grade
Report · partial
Lewis 1.0
Current model
The model used for Lewsearch simulated research studies.
Public evaluations document the study configuration and its scope.
Read evaluationsLewis Lewn 1
Current chatbot model
The current chatbot model in the Lewis family.
A separate public evaluation for this chatbot has not been established.
Read evaluationsEvaluation
Dated comparisons, recorded errors and the limits of each test.
3.72 points MAE
500 historical questions, 50 states and 200 simulated respondents per question. September 1, 2026 run. Calibrated retrospective evaluation; the score is not held out or cross-validated.
Selection note and calculation2.95% MAPE
As of October 19, 2026, three graded spending outcomes and four forecasts. Both NRF forecasts are averaged within their outcome; the three outcomes receive equal weight. The published persistence reference is 2.09% MAPE on the same weights.
Included outcomes and calculationNot yet shown
These headline scores do not establish subgroup validity, future behavior, open-ended text quality or performance on custom customer audiences. We evaluate the question, audience and instrument before interpreting a result.
Lewsearch
Our enterprise product helps research teams study questions with simulated respondents.
Lewsearch for research teamsCompany
Swarmgram, Inc. is the lab behind Lewis and Lewsearch. Jeffrey Mead founded and runs the company.
Read about SwarmgramContact