I appreciate you taking the time to write a detailed reply, and I acknowledge the points you've made. Thank you.
I've written my reply below, but I do feel very passionately about this type of education and would love to continue this discussion elsewhere, if you so like.
I appreciate the 'gap' in testing you speak about in the job market and I entirely agree that separating the programming and mathematical sides of machine learning is a great way to test candidates, but, I feel that ignoring the maths, in this instance, is too big a step.
I disagree with your compiler point because the compiler doesn't generally change the result of your code, but using the wrong machine learning model can. Gradient Descent can lead to local optima, which, if you're unfamiliar with this, can just lead to the wrong set of results. It's why DeepMind are studying into Robust methods of Reinforcement Learning.
Given that point, you say the job market is looking for someone who can complete this task, but can you provide an example of a job where someone would need to build an AI model using something like Tensorflow and where the results are important, but, the individual would not need to know how the underlying tech works? I struggle.
My point is that if you don't care about the maths, then you're completing an engineering task, and therefore you should be tested on engineering, and in this case, this certificate should then be more strongly geared in the other direction.
I 100% understand what you're trying to achieve with this certificate but as tutors we have to hold ourselves to the highest standards and that includes rigorous testing of students and candidates. If the fail rate is 95%, so be it. I know Google employs hundreds of thousands of people and has the smartest people, so maybe I don't know what I'm talking about, but as a practitioner of machine learning - we should endeavour to remain creative and solve the problem in hand: a bad carpenter always blames his tools.
I believe that by adding in some multiple choice questions, you can significantly improve the quality of the exam and better distinguish candidates. You often see this in other (highly related) fields. Questions could be like "which of the following 4 choices is the formulae for an Relu Activation function". Or, "which of the following reasons would you use Genetic Algorithms over Gradient Descent". It doesn't need to be the primary focus, but some effort in this direction makes a lot of sense here.