The AI that isn't allowed to make things up

This is the second half of a pair. In the first piece I wrote about the calm half — the signals Pip reads on the stocks you own, and why the trend is the real story. But all of it rests on one uncomfortable fact: it's an AI doing the reading. And an AI that reads your stocks should make you suspicious.

Here's why. Large language models have a talent that's also a liability: they sound exactly as confident when they're wrong as when they're right. Ask one about a stock you own and you'll get a fluent, plausible, well-punctuated answer — with no way to tell whether it read a real filing or invented one. For a chatbot, that's annoying. For anything touching your money, it's disqualifying.

So Pip is built the other way around, with the AI as the last step, not the first. Here's how.

(Same aside as before: none of this is advice or a prediction. Pip describes what happened and sources it; the decision is always yours.)

More signals isn't always more sure

Start with a mistake that would be easy to make: count the signals, and the more that light up, the surer you are.

It's wrong — and wrong in a way that quietly inflates confidence exactly when you shouldn't trust it. Say a company files an 8-K and the stock drops 4% the same morning. That looks like two signals agreeing. It isn't. The price moved because of the news — it's an echo, not a second, independent witness. Counting it as confirmation is like asking someone to repeat what they said and calling it a second opinion.

So Pip weighs signals by their independence. A filing plus a price move that just reflects it is one story, well-sourced — not two. But a filing, plus unrelated insider activity, plus a real shift in analyst expectations — three things that didn't cause each other — is genuine corroboration, and confidence rises accordingly. I think of it as echo-humility: the difference between measuring the world and hearing your own voice bounce back.

Confidence, not certainty

So Pip tells you how sure it is — and caps how sure it's allowed to be. Thin evidence, a single unconfirmed source, or signals that contradict each other all pull confidence down, and Pip says so out loud instead of papering over it. A quiet, well-sourced "here's what we know, and it's only one source" beats a confident-sounding guess every time.

There's a materiality question hiding in every move, too, and it isn't one-size-fits-all. A 5% day is a real event for a sleepy utility and a Tuesday for a volatile growth name. Pip judges a move relative to how that specific stock normally behaves — so a wobble that's really just the market breathing doesn't get dressed up as company news.

The part that keeps the AI honest

Now the design choice everything rests on.

In Pip, the language model does not decide what is true. It doesn't judge what moved, whether it's material, whether the sources agree, or how confident to be. All of that happens first, in plain deterministic code — the same way every time, checkable, no vibes. Only after that verified, sourced result exists does the model get involved, and its job is narrow: put it into clear English.

Deterministic code certifies. The model narrates. Nothing else.

That ordering is the whole trick. Let a language model decide what's true and you get a horoscope — fluent, confident, untethered. By the time it reaches the model in Pip, the facts are settled and every one has a source attached; the model can only phrase what's already been checked. So when a briefing says a stock moved on an 8-K, there is an 8-K — and you can tap the link and read it yourself. Every claim traces back to something you can open.

That's how you get the useful part of AI — plain-English summaries of things that would take you an hour to piece together — without the part where it makes things up.

What Pip won't do (on purpose)

Some of the most important engineering in Pip is the stuff it refuses to do. It won't tell you to buy, sell, or hold. It won't set a price target. It won't predict where a stock is going — not because that's hard (though it is), but because that's not the job. Pip's job ends at "here's what happened, here's why it might matter to the reason you gave for owning this, here's the source." The next move is yours.

The point

I didn't want a smarter oracle. I wanted a trustworthy read — sourced, corroborated, honest about what it doesn't know — so I could get caught up on my own portfolio in about two minutes and still make my own decisions. Built for the 90%, as I put it in the first piece: the engine isn't here to make you feel something. It's here to tell you what's true, plainly, so you stay the one in charge of your own head.

If that's what you've wanted for the stocks you own, Pip is opening to a small first group →.