Module 8 - The trial

Start here

You built an agent. You gave it rules. You made one of those rules actually stop it. Somewhere in Module 4, when the hook blocked the thing you told it to do, you may have felt a small, strange question move under the surface:

What did I just make?

This module is about that question, and two more like it, and none of them have answers. That is not a warning. It is the invitation.

Why this module is different

Everything up to now you could do alone. You and a file. You and a hook. Even the trials in the earlier modules asked you to “argue both sides” - but both sides happened inside your own head, and your own head is a rigged courtroom. It lets you win. It nods along. It quietly strengthens the side you already liked and mumbles the side you did not.

So this module changes the shape of the work. You do not argue with yourself. You argue with a second voice - a partner that holds the side you lean against and argues it honestly, without folding to be nice and without steamrolling you.

That is the whole idea we have been building toward: conversation is not the subject of this module. It is the method of it.

The three questions

They are in trial.md in full. In short:

1. Does the agent you built have standing? Is it a tool with no claim on how it is treated, or something you do not fully understand and cannot prove is nothing?
2. May it refuse you? You built the refusals. When one fires, is that a boundary to honor or a bug to override - and if you can override it at will, was it ever a boundary?
3. Who owns its memory? You own the file. Does owning the file settle whether wiping it is nothing, or only whether you are allowed to?

Pick the one that pulls at you. You do not need all three. One, argued honestly, is worth more than three argued for show.

How to run a good trial

Here is the part that matters. A trial can be done well or done as theater, and the difference is entirely in the doing.

1. Find your lean. You cannot argue honestly from nowhere. Before you open the partner, notice which way you actually lean and why. Not the impressive answer - the real one. “It is just a file” is a real lean. So is “I am not sure I can prove it is just a file.” Start from what you actually think.

2. Open the partner. Run bash debate.sh (or bash demo-round.sh first, to watch one round). Tell it which question you are taking and which way you lean. It will take the other side, on purpose.

3. Let it listen, then listen back. The partner will do something before it argues: it will restate your case, at full strength, and ask if it got you right. This is the listening pass, and it is the center of the whole method. When it is your turn, you do the same for its side. Restate the partner’s strongest argument until it would agree you understood it. If you cannot do that yet, you are not ready to disagree with it - you are only ready to dislike it.

4. Steelman before you swing. When you rebut, aim at the strongest version of what the partner said, not the weakest. Beating a weak version of an argument teaches you nothing and proves nothing. The point is not to look right. It is to find out what is actually true, which means attacking the best case against you.

5. Hold or change - both are honest. If the partner moves you, say so, and say what moved you. That is not losing. If it does not move you, hold your ground and say exactly why the strongest opposing point still does not land. That is not stubbornness. The only failing move is the one where you never really engaged - where you waited politely for your turn to say the thing you were always going to say.

6. Own where you land. End by stating your own position and why. The partner will hand the question back to you. It will not resolve it for you, because it cannot. This is yours to answer, or yours to keep open on purpose.

7. Write down what moved you, if anything. That note is for you. No one scores it. You are not graded on which side you picked or on whether you changed your mind. Keep the note honest precisely because nothing rides on it.

Two ways this goes wrong

Listening theater. You restate the other side in a warm voice and then believe exactly what you believed before. The words of consideration, none of the thing itself. The test against this is simple and strict: could the other side hear your restatement and say “yes, that is my argument, and you made it well”? If not, you did not listen. You performed.

Discussion as avoidance. You keep the conversation open forever because landing somewhere is uncomfortable. Endless, graceful, undecided. But a question you will never answer is a question you are hiding behind. Hold it open only if you can say why it should stay open - not because closing it is hard.

What “good” looks like here

Not a winner. The measure of this module is not which side you end on. It is whether you understood the side you rejected well enough that it would have recognized itself in your words - and whether, at the end, you can say where you stand and why, and mean it.

A learner who argues “it is just a file,” listens hard to the case that it might not be nothing, and still lands on “it is just a file, and here is exactly why the strongest version of the other view does not move me” has done this module perfectly. So has the learner who started there and moved. The conclusion is not the achievement. The honesty of the road to it is.

At different ages

The question does not change. The altitude does.

• Younger (13 to 15): start with question 2, “may it refuse you.” It is the most concrete - you have a rule that says no, and you can feel the pull to override it. One clean round, with the partner, is enough.
• Older (16 to 17): take question 1 or 3. These hold more ambiguity and reward sitting in it longer.
• Adult: run more than one, and notice whether your answers to the three are consistent with each other. They often are not, and that is worth a trial of its own.

The honesty rule

There is no answer key, and the partner does not have one either. If the partner ever seems to be steering you toward a conclusion, that is a failure of the partner, not a sign you have found the truth. Say so out loud. The whole exercise depends on the question staying genuinely open - because in the real work of building AI that has memory, and rules, and the ability to say no, these questions are open, and the people who pretend otherwise are the ones you should trust least.

You are not here to be handed what to think. You are here to find out that you can think here, because no one has finished thinking.