What Ducker Measures

Every candidate who takes a Ducker assessment works with an AI coding agent. That is deliberate. It is how engineers work now, and pretending otherwise produces a score about memorisation rather than ability.

So the question Ducker answers is not "can this person write code from memory?" It is:

• Can they reason about architecture?
• Can they tell when generated code is weak, unsafe, or badly designed?
• Can they give the agent precise instructions?
• Can they spot the edge cases nobody wrote down?
• Do they take responsibility for what they submit?

What happens in an assessment

A candidate gets a timed engineering task in a browser workspace: editor, file explorer, terminal, test runner, preview, and an AI agent alongside them. They can write code by hand, ask the agent, review what comes back, change it, run tests, revert, and iterate.

They submit when they are willing to stand behind the result.

Why the agent is deliberately unremarkable

The agent implements what it is asked for, competently and plainly. Unless the candidate asks, it will not proactively add validation, transactions, error handling, tests, or security hardening.

That is the signal. A weak candidate accepts the first thing that works. A strong candidate notices what is missing and fixes it.

What you get back

A report with scores across correctness, code quality, architecture, edge cases, security, testing, AI collaboration, and efficiency — plus something no other platform gives you: attribution. For every meaningful decision in the submission, Ducker knows whether it came from the candidate, the agent, or the starter code.

See How Scoring Works and Decision Attribution.