I wrote a protocol earlier for testing the model: six glasses, measuring how much of tasting actually lives in language. This piece turns it around. Here the model is the examiner and you are the one who has to back something up.
That sounds unpleasant and it is. It is also the fastest way to notice you use words you have never filled in yourself.
What you want
Taste blind, write down what you taste, then have every term attacked. Not to be proved right, but to see which terms you use because you taste them and which because they belong to the kind of wine you think you are drinking.
The prompt
Below is my tasting note. I tasted blind and do not know what it is.
Attack every term I use. For each one ask: what did I notice that from,
and would I have written it if someone had told me it was a different
grape or a different country?
Flag the terms that come more from expectation than from observation.
Do not guess what it is and do not tell me whether my note is good.
Only ask questions, one at a time.Those last two lines are the core. The moment the model guesses the wine, the conversation is about the wine and no longer about you. And the moment it says your note is good, you have stopped learning.
Taken apart
Forbid the verdict. A model allowed to judge always judges, and usually favourably. That feels nice and produces nothing. Questions only is stricter than any score.
Ask for the counterfactual. “Would you have written this if it came from another country” is the sharpest question there is, because it peels expectation off observation. The word mineral rarely survives that question.
One question at a time. Same rule as the interview for your taste profile. Five questions in one block get answered carelessly, and here the value sits precisely in the discomfort of being pushed.
Where it goes wrong
Three things slip through, and all three are recognisable from the shape of the answer.
It guesses anyway. Halfway through comes “this sounds like a cool-climate riesling”, and from then on the conversation is about that wine instead of about your words. Cut it off and repeat the rule; it happens more as your note gets more specific.
It compliments anyway. “Good observation” is not a question. The moment that creeps in you are done learning, because a model that confirms you costs nothing and returns nothing.
Five questions in one block. Exactly as in the taste-profile interview. You answer all five carelessly, and the discomfort of being pushed one at a time is where the value sits.
Ethics note
A model attacking your words has no idea what is in your glass. So it can only attack what is linguistically weak, which is not the same as what is factually wrong. A term you justify well can still be nonsense, and a term you phrase badly can be exactly right.
This is therefore an exercise in articulation, not in tasting. Mistake it for a tasting exam and you will learn to write more neatly rather than taste more accurately.
And a third: pushing someone on their sensory impressions is intimate work. A model does it with no restraint at all and without noticing you are getting annoyed. That makes it useful, and it is also exactly why it goes differently with people. Stop when it turns into whether your palate is any good, because that is not what this is about.
Take it
Below is what I wrote about [what you assessed].
I did not know beforehand what it was.
Attack every term I use. For each one ask what I noticed it from, and
whether I would have written it for a different [brand / origin].
Flag what comes more from expectation than from observation.
Do not guess what it was and do not tell me whether the note is good.
Only ask questions, one at a time.A small exercise for tonight: write three lines about what you are drinking without looking at the bottle, then cross out every word you could have written without tasting. What is left is your real note.