The Numbers Aren't Up for a Vote

For about a year I described Wolf You Feed as “Palantir in your pocket.” It got a nod at dinner parties. It also got a few worried looks, and it took me a while to figure out why. Half the people I said it to pictured a surveillance dashboard with their name on it, which is close to the opposite of what we built.

The better description, lately, is a quant and a telemedicine doc in your pocket. The quant does your money math before anyone gets to have an opinion about it. The doc pulls the guideline off the shelf and shows you where your number sits. Neither one gets to make things up, and that turns out to be the whole trick.

Some background first. When you bring WYF a hard decision it doesn’t ask one model. It convenes a few of them, from different companies, and lets them argue. Then a separate model arbitrates under a fixed constitution we call the First Law. I still think that’s the right design. Models disagree in useful ways, and a decision that survived an argument is usually better than one that never had to.

Arguments have a failure mode, though. Give three language models a dollar figure and a horizon and ask what you can afford, and you can get three different answers, each one delivered with total confidence. They aren’t lying, exactly. They’re doing arithmetic the way a language model does arithmetic, which means predicting what an answer ought to look like.

So we took the math away from them.

Decide, the default, just gives you the decision. Turn on Precompute and the numbers you gave us go to plain code before the council convenes. Say the question is whether you can carry a $1,100 monthly cost out of $9,000 in savings. A cash-flow projection runs (or a break-even or discounted-cash-flow calculation, if that’s the question you’re really asking) and the result lands in front of the models as a fact they can cite and can’t change. They can still argue about what you should do. They can’t argue about the month the money runs out. If you’d rather decide first and check later, Validate runs the same kind of calculation against a decision you already have and puts the result underneath it, along with the method it used and the formula, so you can see exactly what went in.

Health got the same idea, pointed somewhere else.

A while back I pasted a spreadsheet of blood work into a chatbot and asked it to predict a date of death, give or take. It gave me a range. It also, to its credit, told me the question was a bad one. I’ve thought about both halves of that answer a lot since. The range was a guess wearing a lab coat. The refusal was the smartest thing it said.

Here’s what WYF does instead. Mention that your A1C came back at 6.4, or that your blood pressure was 142 over 91, and it doesn’t speculate about what those numbers mean. It finds the line a medical society actually published – the American Diabetes Association’s Standards of Care, the AHA and ACC blood pressure guideline – and shows you a card with your number next to that line and a link to the source. Name a medication you take and you get the official label from the National Library of Medicine’s DailyMed. The models read those cards the way they read the math. They’re facts, and nobody on the council gets to rewrite them.

A card is a comparison. It says 6.4 sits inside the band ADA calls prediabetes. What that means for you is your physician’s call, and the card says so in plain words. I’d rather be precise and a little boring there than helpful and wrong.

We added one real calculation on the health side this week too: the ADA’s own Type 2 Diabetes Risk Test, the seven-question score the association publishes for the public. You confirm the answers, code adds up the points exactly the way ADA does, and you see whether you’re past the line where ADA suggests talking to someone about getting tested. Ten-year heart risk is next on that list, once a lawyer has signed off on how we show it.

The part I’m proudest of is small and easy to miss. After a decision renders, WYF goes looking for primary sources on whatever topics your numbers touched and shows them to you as candidates. They arrive after the decision, so the decision you just got stands on its own. Open a link. If the source says what the line under it says, admit it, and the next time that topic comes up it’s part of your record. If you’re a paramedic or a nurse you can say so when you admit it, and the record keeps your role exactly as you stated it.

We could have built this the other way around: stop the pipeline, fetch sources, wait for a person to approve them, then decide. It looks cleaner on a whiteboard. It’s also no use to a first responder in a parking lot at two in the morning, and it would double the wait for everyone else. The answer you need now shouldn’t sit in a queue while somebody reads a PDF.

None of this makes the models smarter, and that’s sort of the point. Judgment still lives with the models – the weighing of things that don’t reduce to a number. What we’ve done is shrink the room they have to improvise in. Arithmetic is code, and so are published thresholds. Which sources count is up to you. Whatever is left over is the argument, which is where we wanted the models spending their effort in the first place.

If you’ve got a decision that involves both your body and your bank account (most of the hard ones do), try it with Precompute on. Give it the real numbers. Then watch what the council does when it can’t fudge them.