Records spread across dozens of tables, joined on the values they share

Evidence work rarely lives in one table. A single report becomes a table of sources, a table of indicators, a table of places, a table of findings and a table of quotes, joined by IDs that sit in different columns with different names. Some cells hold one ID; others hold a list. Dates are written three different ways.

The joins are what make the evidence traceable, but they are hard to see. Before you build reports on a set of tables, you need to know which columns really link, which links are noise, and which records hang off nothing.

Read whatever people actually have. CSV, TSV, Excel (every sheet becomes a table), JSON, and zip files holding any of these. Files inside a zip that are not tables, such as PDFs and readme files, are left out and counted.

Profiled every column. Each column is typed as a date, a number, a short label or long text. Only dates, short labels and whole-number columns named like an ID or code can link; decimals and long text never do. Dates written in different formats are normalised to one.

Linked columns on evidence, not names alone. Two columns in different tables link if they share at least two values and at least 30% of the smaller set of values, or if they have the same name and share at least one value. Cells that hold a list of IDs are split, so a record links to every ID in its list.

Switched off the noise, and said why. Status and category columns, catch-all ID lists and columns where every record holds the same value start switched off, with the reason shown. The user can turn any of them back on, switch off any link, or link two columns by hand.

Kept the files in the browser. Files are read in the browser tab; nothing is uploaded or stored. With a visitor's consent, analytics count visits and which features are used. Nothing about the files is sent: no file names, column names or values.