An answer you can check

There is a specific feeling you get when an AI tells you something about a person. It sounds right. It is fluent, confident, appropriately hedged. And you have no idea whether any of it came from that person or from the shape of a thousand people like them.

That feeling is the thing this product has to beat. Not the writing quality, which models won years ago. The checking.

What an answer is made of

When your page answers, three things come back. The answer itself, in your voice. The sentences in it that rest on something you wrote, marked. And underneath, the things you actually wrote, numbered, in your words, not paraphrased.

So a recruiter asking how you think about hiring does not get a plausible essay on hiring. They get a few sentences that quote you, with the two lessons those sentences came from sitting right underneath, dated, in the words you typed. If the answer sounds like you, they can check that it is you, in about four seconds, without leaving the page.

The model is only allowed to cite what it actually used. Padding the list with everything that got retrieved would be the easy way to look thorough, and it would quietly turn the sources into decoration. A list that always looks impressive tells you nothing. A list of two, where both are load bearing, tells you everything.

Why quoting matters more than summarising

A summary is a small rewrite, and a rewrite of what you wrote is not what you wrote.

This matters most on the things you were careful about. If you spent an afternoon getting one sentence right about when you would fire someone, a model paraphrasing that sentence will make it blunter or softer, and either way it will now be a position you did not take, attributed to you, to someone deciding whether to work with you.

So the answer quotes you directly wherever it can, and the sources under it are never paraphrased at all. The model's job is to pick which of your words bear on the question and to join them up. It is not to improve them.

The part that is harder than it sounds

For a while the page would sometimes give a fully sourced answer and then flag it, internally, as an answer it could not give. A model marking its own homework gets it wrong now and then.

The first version of the code took that flag at its word and stripped the sources off. Which produced the one thing this product must never produce: a confident unsourced paragraph about a person, sitting on their own page, in their name.

So the rule now is that an answer keeps what it cited whatever it claims about itself, and the page labels what it is doing instead. When it has a real answer, you get the answer and its sources. When it does not, it says so, and anything it points at is labelled as nearby rather than as your view. Those are different things and they are never allowed to look the same.

Nothing behind it moves

The other half of checkable is that your writing does not get edited under you. Nothing here is updated in place. When you change your mind about something, the old version is not overwritten, it is retired and the new one is written next to it with its date. Your page answers from what is current, and the history is still there, for you.

That sounds like a database detail and it is really an honesty one. A page that quietly rewrites your past positions is a page whose quotes mean nothing. If somebody saw an answer from you in March and a different one in September, both were true when they were given, and you can show that.

It also means you can read back everything your page has ever said in your name, with who asked and what it released. Not a summary of it. The log.

What to do with this

If you are deciding whether to have a page, the thing to test is not whether the answers sound good. They will. Ask it something you know the answer to and look underneath. Either there are your words there, or there are not.

Ask a page and check one. Then, if you want one of your own, make it.