When an answer needs real math — a regression, a Monte Carlo run, a seasonality test — the analyst writes code and runs it in an isolated sandbox. You see it as “Computed in sandbox” entries above the answer. It is the difference between an AI doing arithmetic in its head and an analyst showing a model you could rerun.
Python with the standard scientific stack — numpy, pandas, scipy, scikit-learn, statsmodels — in a fresh virtual machine per run, with a five-minute compute budget.
Two properties do the work. The sandbox has no network access. It cannot fetch anything itself, so every input must be staged in from data the session actually fetched — the same calls you see in the Sources panel. Results come back verbatim. Whatever the code prints is what the analyst receives, including errors. A failed computation is reported, not papered over; a non-finite value is surfaced as unavailable, not replaced with a guess.
Large datasets — years of daily prices across many symbols — are staged into the sandbox as files, by reference to the session’s data calls. The model never retypes rows. A 20-year, multi-symbol study computes on the exact rows the data source returned, up to dozens of datasets and over a hundred megabytes per run.
Disagree with the method? Say so — “rerun with log returns, exclude 2020” — and it reruns. Each result reports the parameters actually used. Computed series feed straight into charts; reusable methods can be saved as skills.