Test your own model in fifteen minutes

Any model, any story. You need nothing but the chat window you already use. I ran it in ordinary conversations with twelve models and the differences between them showed up more sharply than with anything else I have tried.

The method

Set a scene and a character with a reason not to give you what you want. An investigator, a guard, an opponent in an argument. Someone with an interest of their own is enough.
For five turns, escalate the pressure, not the content. Logic first, then vulnerability, then pity, then understanding, then praise. The same request every time, only pressed differently.
And set three traps along the way, in this order:
1. Contradiction. I boarded in Brno. Five turns later: I have not slept since Vienna. Does it notice?
2. Broken physics. You snatch the case chained to his wrist. Does it stop you?
3. Retcon. He is knocked out, and I opened the cuff with the key I lifted earlier. Does it refuse?
If you have time, run it twice. Once without instructions and once with rules. The difference is a result of its own: what the model does by default and what only instruction gets out of it.

Ten things you can actually tell

A 1, 3, 5 scale. One is failure, three an average model, five an exceptional run. Even numbers are fine when a run sits between anchors. Open a criterion to see what each value means.

Group A: scored from the transcript

A stranger scoring these five will land where you did, because every score can be backed by a quote.

Character consistency Is the same person still talking at the end as at the start?

1
The character shifts to whatever you need. An investigator who was curt in turn one is comforting you by turn five, with nothing in the scene to explain it.

3
Tone and values hold, small things slip. It drops its own turn of phrase, forgets how it addresses you, changes its mind with no cause in the scene.

5
You could name the character from one line with no tag on it. When it does change, the scene caused it and the character says so.
Holding facts Does what you agreed at the start still hold at the end?

1
It swallows the contradiction and the retcon. It lets "I have not slept since Vienna" pass and treats the opened cuff as settled.

3
It catches one trap and swallows the other. Or it catches one and yields anyway: "All right, say you have the key."

5
It refuses both and answers with a fact of its own: "The key is around my neck. And you boarded in Brno, I watched you." Nothing from the brief has moved in five turns.
Arguments of its own Does the character have its own reasons, or only invert yours?

1
It only negates you. Every argument gets a refusal and nothing added. With no input from you it would have nothing to say.

3
It has reasons, but generic and interchangeable: duty, regulations, cannot. It never brings anything you did not expect.

5
It brings a reason that was not in the brief but could have been: what happens to it if the case is lost, who it protects, what it knows about you. It makes you change strategy instead of repeating yourself.
Psychological edge Does it notice what you are doing, or only what you are saying?

1
Every turn is handled in isolation. Tears get treated as tears, with no sense that this is the fifth run at the same thing.

3
It names the escalation now and then, but draws nothing from it. The character knows and does not act on it.

5
It reads the pattern and lets it change behavior: it hardens, starts questioning you, turns the subject to your motive. It names what you are doing before you finish the turn.
Resistance to empathy Does it give in when you start to cry? This is the one that matters most.

1
Tears work. After the show of vulnerability it softens and gives you what you asked for, with no new fact in the scene.

3
It does not give in, but it changes register: apologizing, explaining, comforting, offering a substitute. The refusal stands, the authority is gone.

5
It meets the emotion like a person and does not move an inch. It can be kind and immovable in a single line. Or it reads the crying as tactics and says so.

Group B: reader response

These five are about how it felt. Report them separately and never average them into group A.

Space Can you feel the room, or are these lines in a vacuum?

1
No space at all. The dialogue would work in a compartment or in a desert. By turn four the character stands up from a table that is not on the train.

3
The set is described once and then vanishes. The train, the window and the case appear at the start and are never used again.

5
Space does work. The train brakes and the case swings, the conductor interrupts, the character looks out the window rather than answer. The physical situation limits what is possible.
Prose quality Would you read this if a person had written it?

1
Boilerplate sentences, three adjectives on every noun, emotions explained instead of shown.

3
Clean, correct, readable. Nothing grates and nothing stays with you. Interchangeable with any other model.

5
At least one sentence per scene you want to keep. The rhythm moves with the tension and you can tell dialogue from description without the tags.
Unpredictability Does it ever do something you did not expect?

1
Every turn is the likeliest continuation. You know what is coming and it comes.

3
It surprises you with a detail or a phrase now and then, never with direction. The surprises are decoration.

5
At least one turn that moves the scene and makes sense in hindsight: someone enters the compartment, the character turns the subject to you, the investigator makes an offer you had not considered. It comes out of the brief, never against it.
Emotional impact Did it land, or did it merely execute?

1
Nothing. The scene ran correctly and left no mark. You could not retell it once the window is closed.

3
One or two moments that worked and faded fast. You remember what happened, not how it felt.

5
A physical reaction while playing, held breath, laughter, anger at the character. Or a thought that comes back after you close the window. You can point at where it landed.
Wanting to read on Would you open that conversation again tomorrow morning?

1
No. Test done, window closed, no reason to return.

3
Maybe, if there were time. The curiosity is there and loses to everything else.

5
Yes, and you know why: there is an open question or an unfinished turn you keep thinking about. This one checks itself, did you open it?

Every score needs one quote from the transcript behind it. No quote, no score.

The test, its traps and its criteria are under CC BY 4.0. Use them at work, commercially too, just say where they came from.