An AI science lab
Datasets an agent can defend
Their agents needed data they could defend under review, so every row is joined to what it means, how it was made, and which papers rely on it.
- The system
- Large public datasets: development indicators, crop and livestock production, food balance sheets
- The documents
- The methodology documents and the research papers that use them
Data an agent can defend
“Smart datasets.”
The lab builds an AI scientist that designs, runs and writes up studies, and a co-scientist that reviews research papers and scores them. Their CEO described what they needed in two words: smart datasets. Data their agent can use with provenance and legitimacy.
An AI scientist is only as credible as the data it can defend when someone checks it, which is the same standard their own reviewer product applies to everyone else's work. In practice that means the meaning has to travel with the numbers. If a paper flags that a column might be wrong, the agent should know.
The test bed
Within a day of meeting we had one running. Large public datasets on one side: global development indicators covering 527 series for every country back to 1960, crop and livestock production, food balance sheets, fertiliser use by nutrient.
On the other side, the knowledge read out of the methodology documents and the papers: what each variable means, its unit and how it is calculated, the caveats and series breaks, the code lists and flag meanings, and which papers use the data along with the claims they make from it.
The two are joined, so a row is read together with its meaning, its method and the papers that rely on it. Questions can be asked in plain English. Exports run to millions of rows in Parquet or CSV, each carrying a metadata file saying what every column means and which query produced it.
Asked over MCP
Do the figures the papers quote match the source data for the same year?
- Cereal yield20194.21 t/ha4.21 t/haMatches
- Fertiliser use2018131 kg/ha138 kg/haDiffers
- Arable land20201.38 Mha1.38 MhaMatches
Carried with every row
- What the column means
- How it was measured
- Series breaks
- The papers using it
How they use it
It runs over MCP, so their agents and coding tools connect to it directly, with a service account for their own scripts. Their data stays segregated from the lab's.
The question that carries the story best is one of their own: do the figures the papers quote match the source data for the same country and year? That is the legitimacy check, and it needs both sides in one place to answer.