Notes on how understanding actually works — cognition, systems, incentives, and the difference between feeling competent and being competent. Most of what is here started as something that turned out to be wrong.
How these are made. They come out of long working sessions with Claude. I bring the problem, the argument I want tested, and the judgement about what is true; the drafting, the counterarguments, and a good deal of the structure come out of that conversation. Every piece then goes through a review pipeline — a disclosure scan, a transposition from private notes into something a stranger can use, and two adversarial reviews — before I approve it for publication.
I am not claiming I wrote these alone, because I didn't, and a site whose only editorial rule is if a claim cannot be checked, it does not go in cannot start by misrepresenting its own authorship. What is mine is the choice of problem, the standard applied, and the decision to publish. The errors are mine too.
Each piece states what was expected, what happened, and where the argument was wrong. If a claim here cannot be checked, treat it as decoration.
Essays
-
Verification Is the Binding Constraint
Skill does not increase without a cost for being wrong. Frontier labs pay a 6x spread inside one company for work that resists specification, which is a premium on verification rather than on production. One metric detects whether a working day generated any learning at all.