An AI research firm for the social sciences.

What we believe

AI can now do a large share of the analytical work in a social-science study. It can read the data, write and run the code, and draft the findings. What it cannot yet do reliably is produce the same answer twice, notice a weak comparison, or tell cause from correlation without help. In a randomised trial of 288 researchers, teams that let AI lead reproduced a paper on their own 37% of the time, against 94% for human-only teams (RWI). The value is in the checking.

How we work

We start with the estimand

An estimand is the precise quantity a study sets out to measure: the effect of what, on what outcome, for whom, compared with what, over what period. Writing it down first forces the important choices into the open. More on why this matters.

Every number is executed, not written

Figures in our reports come from code that was run, and rerunning it gives the same value. A number that cannot be traced to a run does not go out.

Results are repeated and cross-checked

We run analyses more than once and compare the results, check totals and sources mechanically, and report the uncertainty, including how much a result moved between runs.

Reproducible handover

You receive the report, a technical appendix and the files needed to rerun every result. Your team, or a reviewer, can check the work without us.

Independence

We are paid for the quality of the analysis, not for a particular answer. Findings are reported as they are, including inconvenient ones. We decline work where the data cannot give an honest answer.

Data handling

We work with de-identified or aggregated data wherever the question allows. Data terms are agreed in writing before anything is shared, and data is held only in environments the client approves and only for as long as the work needs it.