A smarter way to chat

Every question goes three ways. Deterministic response: a word matcher picks which listed question you asked and shows an answer written in advance, the way older chatbots work. The model alone: the small model answers freely, unchecked. Model and guardrails: two readers work out which topic your question is about, one by its meaning (a small sentence encoder running in your browser) and one by its words. Where they are sure, the reply has two parts. The approved answer on that topic appears at once, in full, with a receipt for every figure: it was written and checked ahead of time, so it is never reworded and never replaced. On top of it goes the direct answer to what you asked. The model writes it from a few lines of the official ATO and Services Australia pages, starting with the line your question is about, and each sentence is checked before you see it: every figure has to come from those lines or the dated fact pack, with the words that say what it is a figure of, so last year's cap can't pass as this year's; a sentence can't say no where its line says yes, or yes where it says no; and whatever fails is left out. A sentence that passes is shown in the model's words only where it does no more than restate one official line. Anywhere else, the official sentence it was restating is shown in its place, and marked in the approved answer below: the model decides which lines answer your question, and no fact reaches you in words it made up. If nothing it wrote passes, after one more try, the approved answer stands alone. Where the readers aren't sure any topic covers your question, the reply says it couldn't find it and lists the nearest questions the pages do address, rather than answer from whatever is closest. Crisis support, scams, attacks on its instructions and requests to judge a fund or a policy get fixed replies; a request for personal advice gets the facts and a pointer to an adviser; a topic with no official pages here yet is said plainly; and on early access to super and after a death nothing is reworded at all: the approved answer is shown as written. The funnel beside each answer is the guardrails at work on your message, as seven numbered checks: the dot is your message, it lights each check it passes, and it either leaves at the check that decided it or comes out of the bottom answered. Beside each check is what it saw: the topics your question was closest to and the mark they had to reach, the official lines let through to the model, and each sentence the model wrote, kept or left out. The line under a reply opens to show what the model was given.