All six dimensions
Code writing
The baseline. A candidate is given a specification and has to produce working code against it. A grader model reads the result against that specification point by point, so the score reflects which stated requirements were actually met.
What a candidate is asked to do
- 01Implement a function to a written specification.
- 02Complete a partially written module so it satisfies the stated behaviour.
- 03Handle the stated edge cases, not only the happy path.
How it is scored
What earns points
- Meeting every requirement in the spec, including the awkward ones
- Handling empty, boundary, and error inputs as specified
- Code that reads like the surrounding codebase
What loses points
- Solutions that handle the obvious cases and miss the stated edge cases
- Inventing requirements the spec did not ask for
Why it predicts the job
Specifications are the contract between an engineer and everyone downstream. Meeting one exactly is a more useful signal than solving a puzzle quickly.
The other five
AI collaboration
Five sub-tests covering prompt quality, reading AI code, fixing it, critique, and live collaboration.
Code reading
Untangle real code and say what it actually does.
Debugging
Find the bug in messy legacy code.
Communication
Write for the next human, not the compiler.
Tradeoffs
Justify the choice.