Field notes
Put validation inside the workflow
A faster report is useful when its numbers can be checked.
01
Make the target explicit
In the Beer Sheva model-report project, the specification defines how each table should be calculated. The workflow reconstructs those tables from raw model outputs, giving the review a concrete reference rather than relying on how a report looks.
02
Show the difference
The reconstructed tables are compared with the official target. Missing rows, extra rows and values outside tolerance become visible differences. An analyst can inspect a specific table and time period, then trace individual discrepancies.
03
Make it repeatable
The deliverable includes inputs, scripts, an interface and documentation so the customer can run the checks independently. This is a useful principle for AI-assisted work too: define the expected output, expose uncertainty and give the reviewer a clear way to verify the result.