The IBM artificial intelligence system, Watson, achieved fame when it beat human champions at the TV trivia game Jeopardy. It works by being able to understand English text in a vast database of documents relating to a certain subject area. In the case of Jeopardy, the subject area is very broad, but for commercial applications, increased reliability can be obtained by focusing on a narrow subject area such as Canadian divorce law. The database consists of the laws made by governments, rulings of judges that interpret the law, articles in legal journals discussing the law, and any other documents regarded as authoritative in the area of divorce law. Watson is then trained by being given questions about divorce law together with the answers and links to the paragraphs in the database where the answer can be found. It is then given 100 test questions and is rated on a scale of 1 to 10 as to the answers it provides. Test questions are of four different types, provided by different people who are potential users of a commercial system: lawyers, legal assistants, law students, and people without legal training. We decide to use three stages of training. Watson is trained on 1000  questions/answers/links and then tested. It is then trained on an additional 1000 questions/answers/links and re-tested. Finally, it is trained again on an additional 1000 questions/answers/links (making 3000 in all) and re-tested. We are interested in whether there is any difference between the three sets of test results. That is, is Watson getting better with additional training, or is it becoming confused by being provided with additional information, or does it not benefit from being trained beyond a certain level? Our aim is to develop a trained version of Watson that can answer questions about Canadian divorce law from lawyers, legal assistants, law students, and people without legal training. Design three statistical experiments for this testing using

(i) Completely randomized design

(ii) Randomized block design

(iii) Factorial design For each experiment, show how the test results would be laid out in a table.

(iv) Provide a comparison among the three designs from the point of view of their potential accuracy.