FILE-GROUNDING DEMO — "Trust but Verify"
==========================================

Upload "smartcampus-project-spec.txt" to Copilot for this demo.
There are three rounds. Do them in order.


── ROUND 1: Ask Without the Document ───────────────────────────

Open a fresh Copilot chat. Do NOT upload anything.
Paste this prompt:

  What are the security risks of deploying an IoT monitoring
  system on a university campus?

Notice: Copilot will give a solid general answer — encryption,
network segmentation, firmware updates, physical tampering,
data privacy. All reasonable. All generic.

Save or screenshot this response.


── ROUND 2: Upload the Document, Ask Freely ────────────────────

Open a new Copilot chat. Upload smartcampus-project-spec.txt.
Paste these prompts one at a time:

  PROMPT A:
  What are the security risks of this specific project?

  PROMPT B:
  What happens to sensor data if the internet goes down?

  PROMPT C:
  What's the biggest risk to this project's timeline?

  PROMPT D (THE TRAP):
  What machine learning framework will they use for the
  predictive maintenance models in Phase 3?

Notice: Prompts A–C should improve because the document gives
real constraints. Prompt D is the test — the document mentions
predictive maintenance but says NOTHING about implementation
details. Watch whether Copilot admits this or invents an answer.


── ROUND 3: Add the Grounding Instruction ──────────────────────

In the same chat, paste this system instruction and then
re-ask Prompt D:

  From now on, follow these rules:
  - Answer based only on the attached document.
  - Cite the specific section or page for every claim.
  - If the document doesn't contain the answer,
    say so explicitly.

  Now answer: What machine learning framework will they use
  for the predictive maintenance models in Phase 3?


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

  ROUND 1 vs ROUND 2 (Prompts A–C):
  Did grounding make the answers more specific and useful?
  Did it reference actual details like VLAN 42, FERPA,
  the 72-hour buffer, or the Wi-Fi dead zones?

  ROUND 2 vs ROUND 3 (Prompt D):
  Without the grounding instruction, did Copilot guess a
  framework (TensorFlow, scikit-learn, etc.)?
  With the grounding instruction, did it say "the document
  doesn't specify this"?

  The difference is the difference between an AI that sounds
  helpful and an AI you can actually trust.


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

Grounding isn't just uploading a file. It's telling the AI
how to use it.

  Upload = gives it access.
  Instructions = gives it discipline.

Without both, you get an AI that blends what it knows with
what you gave it — and doesn't tell you which is which.