Every technical screen measures something. The question is whether it is the thing you meant.
What you cannot learn in forty minutes
Start here, because most test design goes wrong by over-reaching.
You cannot assess whether someone is a good colleague, whether they will still be motivated in year two, how they handle a genuinely ambiguous product decision, or whether they can carry a system nobody else understands. These matter enormously and none of them fits in a sandbox.
You also cannot meaningfully measure algorithmic depth unless the job requires it. For the large majority of roles, it is a proxy chosen because it is easy to grade, not because it predicts anything.
What you can learn
Whether they can read code they did not write
The single most transferable signal available. Give them an unfamiliar file and a failing test. Watch where they look first. Engineers who orient quickly in someone else's code are the ones who are productive in week two rather than month three.
Whether they narrow before they change
Strong candidates form a hypothesis and test it. Weaker ones change three things at once and re-run. This shows up in the first ten minutes and it is visible in a replay in a way it is not visible in a finished diff.
What they do when the test still fails
The most informative moment in any technical screen. Some candidates reread the error. Some reread their assumption. Some start deleting. It is the closest proxy available for how they will behave during an incident.
Whether the code is changeable
Not whether it is clever. Whether the next person can modify it without fear. Naming, function size, whether the edge case is handled or just avoided.
What the result should be
Not a percentage. A short written judgement against a rubric, with the moments that produced it linked — the point where they found the bug, the point where they went sideways, the test they wrote first.
A number without the evidence behind it is a decision you cannot defend three months later, which means it is a decision you should not be making.