You would continue with the mistake, and see the construction through. There's a few reasons for this:

• Firstly, the exercise itself is less concerned with drawing the object you're referencing accurately, and more about using that reference as a source of information to help you piece together what is essentially a 3D spatial puzzle. The part that's most valuable to us is how you're forced to think through how the forms you're adding sit in 3D space, and how they relate to one another within it. Having to think about this over and over, across many different studies, gradually rewires your brain's subconscious grasp of how those relationships work. So a mistake may cause the end result to turn out looking different from what you intended, but that doesn't actually make the exercise less effective.

• Secondly, correcting mistakes can trick the brain into believing the mistake never occurred on a subconscious level, whereas leaving it as is and having to work with it makes it very difficult to ignore. And therefore we are more likely to learn from that mistake, to consider why it occurred and how we might avoid it in the future (often it's just a matter of taking more time, rushing less).

• Thirdly, drawing another form on top of the old form will severely undermine your own suspension of disbelief - that is, your brain's capacity to believe in the illusion you're creating - and that'll make the exercise as a whole less effective.