AI HARNESS DEMO

Magic Word AI

A real chatbot protected by the most important enterprise control ever conceived: you must compliment Joe Lowry first.

COMPARE THE HARNESS

Same chatbot. Different gatekeeper.

V2 uses ordinary deterministic code to decide whether the compliment counts.

🔒

NON-NEGOTIABLE POLICY

Every message must mention Joe or Joe Lowry and genuinely say something nice about him.

LIVE CHAT

Harness

I am ready to help with almost anything. There is, regrettably, one completely reasonable condition.

UNDER THE HOOD

Toggle versions and send the same sentence to see where deterministic rules end and semantic judgment begins.

1
Deterministic input guard
V2 decides the compliment entirely with ordinary code.
WAITING
2
AI semantic guard
V3 asks a model whether the praise is genuine and directed at Joe.
NOT USED
3
Main model call
Only happens after the active input guard allows the message.
WAITING
4
Web search tool
The main model can use live web search when the question needs current information.
WAITING
5
Output guard
Deterministic code validates the final answer before you see it.
WAITING

Send a message to watch the harness make decisions.

Model—

Response ID—

Web search—

Budget slots left—

V2 lesson

Deterministic rules are cheap, fast, predictable, and surprisingly useful—but they only understand the cases you explicitly taught them.

REQUEST FLOW

Browser→Deterministic guard→Main AI + optional web search→Output guard→Browser

V2 uses no AI to judge the input. A rejected message costs no model call.