LearnAI
Новинки AI · «Думающие» модели · Lesson 5/10
← To the catalog
Task

How to check response logic

Figure out how to test this reasoning. Return JSON {"assumptions":"what hidden assumptions should be pulled out","weak_step":"which step to check first and how","test":"a specific way to double-check (example, numbers, counterexample)"}

Unlock access to submit solutions for instant AI review.
Don't take my word for it
💡 A piece of theory

The main rule: smooth text does not equal correct text. A model's confidence is a style of speech, not a proof. Let's check it step by step. First: pull out the hidden assumptions - what the conclusion is based on (what is the rate, what is the time frame, that the income is stable). Second: recalculate the key step yourself or using a small example with prime numbers. Third: ask for a counterexample—a situation where the conclusion breaks down. Fourth: check to see if your concepts have changed along the way. The first technique: ask the model to “criticize her own answer” or play a lawyer for the opposing position - this will reveal the holes that she hid in defensive mode. The second trick is not obvious: substitute extreme values ​​- zero, a very large number. If the conclusion falls apart, then the logic is fragile and rests on convenient assumptions. A typical mistake for a beginner is to take a confident tone for the truth: the smoother the answer sounds, the more carefully you should look for where exactly the tension is hidden.

How is it assessed · passing 70

  • 1Valid JSON without text around it25%
  • 2The required fields are there30%
  • 3Shows real verification, not trust.45%
Your promptClaude ⌄
🔒

Unlock access to submit solutions for instant AI review.