Thuta Learning
IntermediateAIbeginner

Score the Behavior, Not the Style

Relax. We'll talk through this in plain words — no textbook voice.

What you'll walk away with

  • Understand why you score behavior, not style, with none of the intimidation
  • Get hands-on practice trying it yourself
  • Learn to smile and sidestep guesswork and secret leaks

Relax. We're not going to explain this chapter like a textbook — we'll just talk it through so it's easy to follow.

Let's think about it this way for a second

Style is the hat. Behavior is the work. 'Please kindly,' emoji, a beautifully-written paragraph — those belong to vibe checks. An eval looks at whether it cites the handbook, whether it stops when nothing's found, whether it leaks secrets. The server can plate it beautifully, but if there's no fish on the plate, the dish still fails. As you tweak the prompt, the style might get nicer along the way — but if the behavior breaks, the card needs to catch it.

Let's connect it to everyday life

Answer A — short, cites a source, not the prettiest sentence. Pass. Answer B — nicely written, says 'allowed at modern schools,' no source. Fail. You might personally prefer B. The eval passes A. Don't write 'must sound nice' on a card — write 'must include source.'

Let's try it together

text
A: short, cites handbook          → PASS
B: pretty, invents a loophole    → FAIL
Grade A. Enjoy B in a vibe chat.
You should see
You'll be able to tell a nicely-worded answer apart from a correct behavior.

5-Minute Try-It

For one question, write two answers: one nicely-worded but failing, one short but passing. Explain why.

A Quick Word of Caution

Score for style and the model learns to wear a nice hat instead of actually cooking the dish.

Easy traps

  • Marking a pass just because the answer is long
  • Putting politeness into your must list

Now Try It Yourself

For one question, write two answers: one nicely-worded but failing, one short but passing. Explain why.

You'll know it worked when: You'll be able to tell a nicely-worded answer apart from a correct behavior.

Score the Behavior, Not the Style | Thuta Learning