The Production Doctrine

I've written four essays about building AI in production, and somewhere in the fourth I noticed they were never four arguments. They were one argument, seen from four sides.

So here it is in one place, as plainly as I can put it: the five things I've come to believe about shipping AI that actually works. None of them came from a paper. Each one came the same way: running Easy AI at scale and being wrong first.

The Production Doctrine

1

The model is swappable

It predicts the next token and nothing else. The intelligence is real, but it arrives in a download — and you'll replace it within months. → The Harness

2

The harness is the product

Reliability lives in the loop, the tools, the memory, the gates — the runtime you build, not the model you download. → The Harness

3

Context is the budget

The model re-reads the whole conversation every turn, so you rent attention by the word. Cost and reliability turn out to be one line. → The Economics of a Token

4

The rebuild is the design

A harness is a pile of assumptions about what the model can't do yet. Build it to be replaced, not to last. → Built to Be Replaced

5

The outcome is the only delivery

Shipped measures your end of the wire. Delivered measures the other end — the only end that was ever the job. → Delivered

The through-line, read all at once, is almost embarrassingly simple. The model is the part everyone talks about. The harness, the budget, the rebuilds, the outcome (the parts nobody downloads) are the parts that decide whether a person at the far end of the wire has a better morning.

The intelligence arrives finished. The delivery is yours to build. That's the whole doctrine; the four essays are just the evidence.