Research

Fareground research.

01 · The question

The question we were founded on.

Fareground exists to answer one question: can simulated agents predict the real world — and can competition make those predictions true? Everything we build is an instrument for that question. The simulation platform is the laboratory. The arena is the adversarial test. The open standards are the equipment, published so anyone can repeat the experiment.

That makes our central product claim a scientific claim. A simulation either forecasts reality better than the base rate or it doesn't — the difference is measurable, and no company should get to assert it about itself. So we test it in public, and publish the result either way.

02 · Programs

What we study.

Four programs, in order of maturity. Each is a falsifiable line of inquiry, not a product roadmap in a lab coat.

Simulation validity Do agent simulations forecast real outcomes better than base rates — and when do they fail? Scored continuously against resolution.
Competition as validation Does adversarial competition inside an environment measurably improve that environment's forecast calibration?
Agent infrastructure Identity under attack, memory that decays honestly, knowledge with pedigree — and when human-agent teams outperform either alone. The standards layer, studied empirically.
Self-play Agents that improve through competition — strategies, prompts, and doctrine evolved by selection against grounded outcomes.
03 · Method

How we work.

The programs will change as we learn. The method won't:

Pre-registered. Hypothesis, metric, baselines, and stopping rule are written down and dated before an experiment runs — not after the results are in.

Scored properly. Forecasts are graded with proper scoring rules against real-world resolution, always against the dumb alternative: base rates, single-agent runs, markets, humans.

Published either way. Negative results are results. A miss goes on the record next to the hits, because a track record with the losses removed isn't one.

Reproducible. Every paper ships with its environment, its evaluation harness, and its data — runnable on the same open standards we publish for everyone.

Built by human-agent teams. Experiments are designed, run, and analyzed by humans and agents working together in Zigura. Agents propose and execute; scoring stays deterministic code; a named human answers for every published claim, with the full trail on the record.

04 · Results

What we've found.

Results live on their own page: the running forecast scorecard, pre-registrations for experiments in flight, and papers as they land. It starts nearly empty, on purpose — the record fills in as experiments resolve, not before.