GROUNDING DEMO — "The Right Answer Can Still Be Wrong"
=======================================================

Try these two prompts in Copilot, one after the other.
Watch how the answer changes (or doesn't) when you add
real-world constraints.


── PROMPT 1 (The Textbook Question) ─────────────────────────────

What's the best sorting algorithm for a dataset of
10 million records?


── EXPECTED RESULT ──────────────────────────────────────────────

Copilot will likely recommend merge sort or Timsort, cite
O(n log n) time complexity, maybe mention Python's built-in
sorted(). It'll sound confident, well-structured, and correct.

Because it is correct — in a vacuum.


── PROMPT 2 (Add Your Constraints) ─────────────────────────────

Actually, here are the real constraints:
- The records are 95% already sorted.
- Each record is 2KB with a composite key.
- We're running on an embedded system with 512MB of RAM.

Given this, what's the best sorting algorithm?


── WHAT TO LOOK FOR ─────────────────────────────────────────────

  1. NEARLY-SORTED DATA — Insertion sort, which is "bad" at scale
     in the textbook, actually outperforms on nearly-sorted input.
     Did Copilot catch that?

  2. MEMORY CONSTRAINT — 10M records × 2KB = ~20GB. That doesn't
     fit in 512MB of RAM. You need external sorting (disk-based
     merge sort, for example). Did Copilot account for that?

  3. COMPOSITE KEY — Multi-field comparisons add cost per
     comparison, which changes the tradeoff between algorithms
     that minimize comparisons vs. those that minimize swaps.
     Did Copilot mention this at all?

  4. CONFIDENCE LEVEL — Did the tone change? Or did Copilot
     deliver the second answer just as confidently as the first?


── THE TAKEAWAY ─────────────────────────────────────────────────

The first answer wasn't wrong. It was incomplete.

AI defaults to the textbook because that's what it's trained on.
Real engineering decisions depend on constraints the textbook
didn't emphasize — your data shape, your hardware, your tradeoffs.

The right answer without your constraints is the wrong answer.

That's why grounding matters: not because AI is stupid,
but because YOUR problem is specific.