The Eval Is in Your Stomach

Two words

I give my students a two-word prompt: order dinner. Then I ask how they would know it worked. Most of them say the order was confirmed. I tell them the eval is the food in your stomach. Not ordered. Delivered, brought inside, put on the table, eaten.

They laugh, and then they get quiet, because they see how much sits between those two finish lines.

The principle is older than any of this. A task is finished where the want is met, not where the doer stops. Telephone engineers gave the problem its name a long time ago: the last mile. The long haul across the country was the easy part. The short run from the pole to the kitchen wall was where the cost and the failures lived. Every delivery business since has relearned it.

Ordered is a status. Fed is an outcome.

Where the labs stop

Earlier today I wrote about a new benchmark for personal AI. It grades an assistant on the errands of a life, and it does something right: it gives credit only for what was completed, not for a good-sounding reply. But completed, there, means the order exists in the store’s system. The researchers imagined an assistant that could do no physical work, so they graded the last thing it could touch.

That is fair, and it is not enough. Nobody ever wanted an order.

It starts before the order

The right eval reaches backward as well as forward. Suppose the skill your AI works from says you are the head of a household of four. Then order dinner means four stomachs, unless one of them is at a sleepover tonight. An AI graded on the order buys four meals and passes. An AI graded on the table checks the family calendar first, sees who is home, and buys three.

Whose stomachs is part of done.

Four handoffs

Between the order and the table there are four places where dinner fails.

1. Placed to accepted.

   The kitchen is closed, or one dish is out. The order exists and the dinner does not.

2. Accepted to on the way.

   The bag sits on a counter for forty minutes while the status says preparing.

3. On the way to the door.

   The gate code was never passed along. The bag lands on a neighbor’s step.

4. The door to the table.

   It arrived. Nobody knew. It is cooling on the porch.

An AI held to the stomach watches every one of these. It puts the gate code in the instructions. It notices the order is late and calls. It reorders the missing dish. It cannot carry the bag, but it can own the outcome, and owning the outcome is different from performing every step.

It summons you

Here is the part people do not expect. The AI is tracking the driver. It knows the arrival time better than anyone in the house. So two minutes out, it calls a human to the door.

For half a minute the direction of instruction reverses. You asked it for dinner, and now it is telling you where to stand. That is not the machine taking over. It is what any good manager does with the one step they cannot do themselves: they put the right person in the right place at the right time, and they say why.

Then it does the last thing a screen can never do alone. It asks. An AI cannot see a stomach, so the task stays open until a person says it is done. Silence is not a yes.

What I expect

1. “Delivered” stops meaning done.

   Apps grow one more state after delivered, and only the person who wanted the thing can set it.

2. Assistants are judged by their closing question.

   An AI that never asks how it turned out will feel careless, the way a contractor who never comes back to look feels careless.

3. Being summoned by your own AI becomes ordinary.

   “Be at the door at 7:02” will feel no stranger than a calendar reminder, except that it comes with a reason.

4. Businesses compete on the last handoff.

   The ones that show your AI live status all the way to the doorstep win. The ones with a dark stretch between shipped and arrived lose customers they never hear from.

5. The benchmarks follow the food.

   The next round of tests for personal AI will include a person, a few days later, saying whether it worked.

I teach this with dinner because nobody argues with dinner. But I hold my own AI to the same rule. When I hand something off, the work has to get done, not get started, and I say what done means out in the world, not on the screen. At the lab in Charleston it is becoming the first line at the top of a person’s skill: done means. Everything under it gets easier.

The claim, and where it breaks

Already true

An AI today can check a calendar, place an order, track a driver, send a text and ask a question. Every piece of the stomach eval already exists. What is missing is the instruction to keep going.

What has to happen

People have to write “done means” in their own words. Businesses have to let a customer’s AI see status past the sale. Assistants have to be allowed to stay on a task for hours, not seconds.

Where I am probably wrong

People may find the closing question tiresome. If every errand ends with “did it work?”, we will have rebuilt the survey, and people will stop answering. The AI has to learn when asking is worth the interruption. If it cannot, this standard is right and nobody will keep it.

Back on Thursday, the photograph is still on your phone. An AI graded at the order finished its work twenty minutes ago and has no idea anything is wrong. An AI graded in your stomach saw the wrong doormat the moment the picture arrived, messaged the driver, and sent someone next door.

Done is not on the screen. It is on the table.