Field note · October 20, 2026
The skill we should have taught: verifying work you did not write
Reviewing a query, a model, or an analysis you did not author is now a bigger part of the job than authoring it, and we have almost nothing on it.
The biggest gap in this course, named plainly: it teaches you to write things and barely teaches you to check things.
That ratio no longer matches the job. A large share of the SQL, the transformations and the analysis code a person encounters in a week was written by someone or something else. The bottleneck has moved from production to verification.
We are drafting a module on it. The shape so far:
Verify the grain before the logic. Before reading a single line of a query's logic, establish what one row of its output is meant to be, and check the row count against that claim. Most errors are visible at this stage without understanding anything else.
Check the joins independently. For each join, confirm which side is one and which is many, using the data rather than the author's intent. group by key having count(*) > 1 on the supposedly-unique side takes ten seconds and catches the most damaging category of bug.
Reconstruct one number by hand. Pick a single row of the output and recompute it from source, manually. This is slow, feels beneath you, and catches things nothing else does. It is the single highest-yield verification technique and almost nobody does it.
Look for the silent default. Every else, every coalesce, every fillna. Ask what falls into each and whether that is intended. Wrong numbers hide in defaults, because a default never errors.
Check the filters against the population. where region = 'EMEA' — how many rows does that exclude, and is that the number you expected? A filter that silently matches nothing returns a clean, confident zero.
Ask what the author would have had to know. If a feature required information unavailable at prediction time, or a metric required a definition nobody agreed, the work is wrong regardless of how correct the code is.
None of these are new. What is new is that they need to be the primary skill rather than a review checklist, and taught as such. We had it backwards, and correcting that is next year's work.