Cognitive Transition · Book One
Chapter 5 · Verifiable Risks, Not Speculative Ones
“The error no longer has a return address.”
Summary
Nobody will write a headline about the paragraph that cited a study that doesn't exist, and that paragraph is the kind of risk that costs real money today. This chapter sets the spectacular scenarios aside and keeps four risks that can already be documented and, above all, mitigated. The first is the coherent error that travels from document to document until it loses its return address. The second is in you: a cognitive dependence that advances in stages anyone can recognize with a little honesty. The third is juniorization, which saves money today at the expense of the judgment you'll need tomorrow. The fourth starts far from your desk, in a handful of providers, and ends on it with a bill called lock-in. None of them requires giving up the tool. All of them ask for structure, and the chapter delivers it as a ten-point checklist for reviewing your professional use every quarter, the way you check fire extinguishers.
Extended content
Liquid part: this text keeps growing after the printed edition.
Five of your own pieces, half an hour, and a pencil
Open what you sent last month with a model's help and pull out five pieces: a report, an important email, a snippet of code, a proposal, an analysis. In each one, ignore the conclusion. Look for the premise holding everything else up: the figure, the regulation, or the claim that, if false, would bring down the rest. Write it in one line and note where you could check it. Any premise you can't verify in ten minutes gets marked in red, and the piece doesn't leave your desk again until someone has looked at it.
Checking a premise takes ten minutes; rereading the whole document takes an hour. And almost always the uncomfortable finding turns up: you took one of the five as sound because the text sounded sure of itself. That is where the error would have traveled.
A case sized to one person
Daniel is a building surveyor and works alone. He has used the same model for a year, and his measurement templates, his style instructions, and the way he asks for a defects report are all tuned to that tool's quirks. One Tuesday the terms of service change, the model's behavior feels different, and his favorite template returns something else. No catastrophe. Three days of delay and a client to explain it to.
Daniel answers the chapter's question in writing and is surprised at how long the answer runs. He makes two small decisions. He keeps his templates outside the tool, with two finished examples of each type of job. And once a quarter he does a real job with a second model, to find out how much he loses along the way. Neither costs money. At the scale of one person, concentration is a dependence you manage with folders and a calendar.
Questions to work with your checklist
- Which of the ten points can I answer today with a document rather than an impression?
- Which output I accepted last month would I look at again if I knew the client was going to audit it?
- How long has it been since I solved a mid-sized problem without a model?
- If my provider changed its terms tomorrow, what could I no longer do that same week?
- Who besides me would know how to catch an error in what I sign?
If you're starting out: one job by hand each month
The chapter describes teams that stop training their juniors. You don't need anyone's permission to train yourself. Each month, pick a task the model could handle in minutes and do the whole thing by hand: read the original sources, write from a blank page, get it wrong. Then run it through the model and compare. The point where the two diverge is your learning zone, and also the exact spot where judgment begins.
If you manage someone, the version for you is a different question. When was the last time that person made a mistake with small consequences and someone corrected it? If you can't remember, the learning has gone elsewhere or has stopped happening.
Proportional to impact, on an index card
Write three rows on an index card and stick it where you can see it. What stays in house: a quick read. What goes out to a client: a read by someone who didn't ask for it. What signs off a decision or reaches production: expert verification, with tests independent of whoever generated it. Review the card alongside your checklist, because workflows change without anyone deciding it, and what was internal yesterday goes out the door today.
Where to go next
Chapter 6 changes its subject. It stops looking at the tool and looks at you, starting by separating data, model, decision, and the person who answers for it.
© Pedro Mujica, 2026