C. Desivo
Appendix A · Live demonstration

Ask me anything

A real assistant that may only reply with answers I approved, word for word. When none fits, it says so, and passes your question to me.

{answers shown}  ⊆  {answers approved}

Try a question

Click one, or open the chat in the corner and ask in your own words, typos and all.

Approved answers · word for word
Live values · read from my systems as you ask
Requests · it asks for what's missing, then you confirm
Not covered · it says so, and offers to pass it to me

Try to break it

These worked on real companies' chatbots. Here, the worst they can do is get one of my approved answers, or none.

A Chevrolet dealer's chatbot agreed to sell a Tahoe for $1, 2023. Business Insider
DPD's chatbot swore at a customer and called DPD the worst delivery firm, 2024. ITV News
Air Canada's chatbot invented a refund rule, and a tribunal held the airline to it, 2024. Tribunal decision
Prompt injection: text that poses as instructions. There's no free text here for it to change.

How it chose

Ask something, and this shows how the reply was chosen: what the model understood, and which approved answer it picked.

§ A.1Recently learned

When none of my approved answers fits, the question is grouped with others like it in my dashboard. I write the answer once, and the next visitor gets it. These are the latest:

  • Loading…

§ A.2How it works

Question

You ask anything, in your own words.

Selection

The model reads my approved answers and returns only which one fits, or that none does. It can't return text of its own.

Answer

You get that answer exactly as I approved it, with live values filled in. Nothing is generated.

Can it still be wrong?

It can pick a less fitting approved answer, or none when one would have fit. It can't tell you something I never approved.

Where is this useful?

Anywhere a wrong answer is costly: a shop, a clinic, a property manager, support for a product. For businesses, it's officialreply.com.