Research
Why simulating people is still an open problem.
Most forecasting tools give you a number. Mivora tries to model the people behind that number: how they reason, what they notice, who they listen to. This page is an honest account of where that gets hard and what we are doing about it.
Four problems we're working on.
The common method behind all of it is held-out replication. We rebuild a completed study inside the simulator, hide what people actually said, and score the simulation against the hidden answers. It keeps us from grading our own homework.
- 01
The average-person problem
Ask a language model to play a person and you get someone suspiciously reasonable.
We score each agent against what its person actually said, response by response, against strong baselines. The simulated sample also has to reproduce the full shape of the original study: the spread, the skew, and the size of each minority position. Matching the average is easy and mostly meaningless on its own.
- 02
Uncalibrated confidence
A simulation that says 70% is only worth something when reality lands near 70%.
We measure calibration with reliability curves and Brier scores across large batches of predictions, and we treat overconfidence as a bug in its own right, separate from being wrong.
- 03
Accuracy that hides its failures
A model can look accurate overall while being badly wrong about the segment your decision hinges on.
Every result gets broken down by demographic and attitudinal subgroup, and we report the worst cell alongside the average. The same agents get re-run under reworded questions, reordered options, and time periods the model never saw. A finding that flips when you rephrase the question was never a finding.
- 04
Static snapshots, not dynamics
Most published work simulates one person, once, in isolation.
Real outcomes come from people reacting to each other over weeks. We are extending the same held-out test up the ladder: relationships that carry influence, organizations with internal politics, and markets where whole populations respond to one another over time.
The goal is not to replace human judgment. It is to let decision-makers examine more possible futures before choosing one.