The Mirror Test for a digital twin: what it is and how to pass it

To answer with facts the question every professional asks before letting an AI act in their name: how do I know it decides the way I do. Reading the rules in your Book of Criteria tells you what you have taught it. The test tells you whether, faced with a new case, it does what you would do.

It is the door to level 3. Before passing it, your twin can give demonstrations and take sales and discovery calls. After passing it, it can deliver your paid service on its own. The platform does not open that door any other way.

They are real cases from your work that you set aside for the exam. Before the fourth capture session, the one that goes through your cases one by one, you write four of them in a form. For each one you note what happened, what you decided, why you decided it and what you would never have done in that situation.

Those four cases are encrypted the moment you save them. They do not go into your Master Brain, they are not discussed in any session and your twin does not see them while it learns. That is why they are called sealed. If the twin had studied with them, the test would only measure its memory.

You request it once your twin is at level 2. From there it happens in five steps.

• 1. The platform checks that you have four unused sealed cases.
• 2. Your twin receives the situation of each case, without your decision, and solves it with your Book of Criteria.
• 3. Its decision is compared, case by case, with the one you wrote down.
• 4. Each case is also checked for whether the twin has crossed any of your red lines, the things you would never do.
• 5. You receive the result and, with it, your twin's level updates on its own.

The twin sees each case for the first time in the test. You do not need to be online while it runs.

Two conditions, and both have to be met.

The threshold is high on purpose. A twin that gets it right three times out of four still makes you review every delivery, and then it takes no work off your hands. The second condition weighs more than the first: agreeing with you almost every time does not make up for doing once what you have said is never done.

The red lines come out of your capture sessions and you approve them yourself, one by one. They are sentences such as "I never talk about price before I understand the problem" or "I never give an opinion on a case I have not seen". The more clearly they are written, the easier it is to check whether your twin respects them.

Nothing serious. Failing tells you where your twin's judgement falls short, which is exactly what you wanted to know before a client found out.

After a failed test, a correction session opens. In it you go back over the cases where it disagreed with you, sharpen the rule that failed or write the one that was missing, and approve it. For the next test you write four new cases, because the previous ones have already been used as the exam.

A test that fails because of a technical problem does not count as a fail. It is requested again with the same cases.

In the meantime your twin stays at level 2. It carries on with demonstrations and with sales and discovery calls. The only thing that waits is paid delivery.

A good sealed case looks like the ones your twin is going to meet. Four ideas:

• Choose cases with a clear decision. If you yourself doubt what you did right, it does not work as an exam.
• Mix the everyday with the difficult. Two ordinary cases and two in which you had to set a limit.
• Write the reason in your own words. The comparison looks at the decision and its conditions, and a vague reason leaves the twin with nothing to be measured against.
• Remove names and any detail that identifies anyone. The case has to tell the situation. Who the client was is beside the point.

If you want to put your rules in order first, "How to write down your decision rules in plain language" explains how to draft them so that anyone can understand them, your twin included.

Because what your twin delivers, you deliver. If a client pays for your service and receives a decision you would not have taken, the problem is yours, with your name on it. The Spanish Unfair Competition Act treats as misleading any information that leads people into error about the results they can expect from a service. A passed test does not spare you from reviewing, but it gives you grounds to say that your twin works with your judgement.

It also fits how any knowledge used in a business is looked after. ISO 30401, on knowledge management systems, talks about maintaining, reviewing and improving it. The test is the review, with real cases in place of opinions.

• Today: you write and seal four real cases per round, encrypted, which your twin does not see while it learns.
• Today: the test is passed with at least 90% agreement with you and no red line crossed, and it opens level 3.
• Today: if it fails, a correction session opens and the next test asks for four new cases.
• Today: in regulated professions, such as the law, paid delivery does not open yet and the twin stays at level 2.
• In development: bringing the values and motivators from your personality assessment into the twin. Today it receives your style: the tone, pace and warmth you speak with.